FlippyDora commited on
Commit
ec9f252
·
verified ·
1 Parent(s): 3c5464b

Add runs (part 25)

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +1 -0
  2. scenes/isaac/47331202/20261114T065042Z/input/anchor_marked.png +3 -0
  3. scenes/isaac/47331202/20261114T065042Z/input/anchor_reference.png +3 -0
  4. scenes/isaac/47331202/20261114T065042Z/input/generation_spec.json +66 -0
  5. scenes/isaac/47331202/20261114T065042Z/input/prompt.txt +143 -0
  6. scenes/isaac/47331202/20261114T065042Z/input/scale_anchor.json +26 -0
  7. scenes/isaac/47331202/20261114T065042Z/input/sdk_smoke.py +49 -0
  8. scenes/isaac/47331202/20261114T065042Z/input/self_check.py +228 -0
  9. scenes/isaac/47331202/20261114T065042Z/input/video.mp4 +3 -0
  10. scenes/isaac/47331202/20261114T065042Z/run.json +139 -0
  11. scenes/isaac/47331202/20261114T065042Z/submitted_prompt.txt +143 -0
  12. scenes/isaac/47331202/20261114T083400Z/evaluator_protocol.json +0 -0
  13. scenes/isaac/47331202/20261114T083400Z/generation.jsonl +0 -0
  14. scenes/isaac/47331202/20261114T083400Z/input/anchor_marked.png +3 -0
  15. scenes/isaac/47331202/20261114T083400Z/input/anchor_reference.png +3 -0
  16. scenes/isaac/47331202/20261114T083400Z/input/generation_spec.json +66 -0
  17. scenes/isaac/47331202/20261114T083400Z/input/prompt.txt +143 -0
  18. scenes/isaac/47331202/20261114T083400Z/input/scale_anchor.json +26 -0
  19. scenes/isaac/47331202/20261114T083400Z/input/sdk_smoke.py +49 -0
  20. scenes/isaac/47331202/20261114T083400Z/input/self_check.py +228 -0
  21. scenes/isaac/47331202/20261114T083400Z/input/video.mp4 +3 -0
  22. scenes/isaac/47331202/20261114T083400Z/run.json +139 -0
  23. scenes/isaac/47331202/20261114T083400Z/submitted_prompt.txt +143 -0
  24. scenes/isaac/47331524/20261114T061812Z/evaluator_protocol.json +0 -0
  25. scenes/isaac/47331524/20261114T061812Z/generation.jsonl +0 -0
  26. scenes/isaac/47331524/20261114T061812Z/input/anchor_marked.png +3 -0
  27. scenes/isaac/47331524/20261114T061812Z/input/anchor_reference.png +3 -0
  28. scenes/isaac/47331524/20261114T061812Z/input/generation_spec.json +66 -0
  29. scenes/isaac/47331524/20261114T061812Z/input/prompt.txt +143 -0
  30. scenes/isaac/47331524/20261114T061812Z/input/scale_anchor.json +26 -0
  31. scenes/isaac/47331524/20261114T061812Z/input/sdk_smoke.py +49 -0
  32. scenes/isaac/47331524/20261114T061812Z/input/self_check.py +228 -0
  33. scenes/isaac/47331524/20261114T061812Z/input/video.mp4 +3 -0
  34. scenes/isaac/47331524/20261114T061812Z/run.json +139 -0
  35. scenes/isaac/47331524/20261114T061812Z/submitted_prompt.txt +143 -0
  36. scenes/isaac/47331524/20261114T081855Z/evaluator_protocol.json +0 -0
  37. scenes/isaac/47331524/20261114T081855Z/generation.jsonl +0 -0
  38. scenes/isaac/47331524/20261114T081855Z/input/anchor_marked.png +3 -0
  39. scenes/isaac/47331524/20261114T081855Z/input/anchor_reference.png +3 -0
  40. scenes/isaac/47331524/20261114T081855Z/input/generation_spec.json +66 -0
  41. scenes/isaac/47331524/20261114T081855Z/input/prompt.txt +143 -0
  42. scenes/isaac/47331524/20261114T081855Z/input/scale_anchor.json +26 -0
  43. scenes/isaac/47331524/20261114T081855Z/input/sdk_smoke.py +49 -0
  44. scenes/isaac/47331524/20261114T081855Z/input/self_check.py +228 -0
  45. scenes/isaac/47331524/20261114T081855Z/input/video.mp4 +3 -0
  46. scenes/isaac/47331524/20261114T081855Z/run.json +139 -0
  47. scenes/isaac/47331524/20261114T081855Z/submitted_prompt.txt +143 -0
  48. scenes/isaac/47334756/20261114T061724Z/evaluator_protocol.json +0 -0
  49. scenes/isaac/47334756/20261114T061724Z/generation.jsonl +0 -0
  50. scenes/isaac/47334756/20261114T061724Z/input/anchor_marked.png +3 -0
.gitattributes CHANGED
@@ -5903,3 +5903,4 @@ scenes/isaac/42899474/20261114T064553Z/generation.jsonl filter=lfs diff=lfs merg
5903
  scenes/isaac/44358647/20261114T061621Z/generation.jsonl filter=lfs diff=lfs merge=lfs -text
5904
  scenes/isaac/45662970/20261114T064610Z/generation.jsonl filter=lfs diff=lfs merge=lfs -text
5905
  scenes/isaac/47115300/20261114T062346Z/generation.jsonl filter=lfs diff=lfs merge=lfs -text
 
 
5903
  scenes/isaac/44358647/20261114T061621Z/generation.jsonl filter=lfs diff=lfs merge=lfs -text
5904
  scenes/isaac/45662970/20261114T064610Z/generation.jsonl filter=lfs diff=lfs merge=lfs -text
5905
  scenes/isaac/47115300/20261114T062346Z/generation.jsonl filter=lfs diff=lfs merge=lfs -text
5906
+ scenes/isaac/47429874/20261114T062029Z/generation.jsonl filter=lfs diff=lfs merge=lfs -text
scenes/isaac/47331202/20261114T065042Z/input/anchor_marked.png ADDED

Git LFS Details

  • SHA256: f4710332f02e78fd95d56162c4edb4dc504ae29bb4dc9bce768e4cfc177b446c
  • Pointer size: 130 Bytes
  • Size of remote file: 73.1 kB
scenes/isaac/47331202/20261114T065042Z/input/anchor_reference.png ADDED

Git LFS Details

  • SHA256: bddaaec0975f8fa3841f081c6864a05cbc041b0e56aa8c5bcf706257c83c7bb3
  • Pointer size: 130 Bytes
  • Size of remote file: 74.4 kB
scenes/isaac/47331202/20261114T065042Z/input/generation_spec.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "contract": "scenes-isaac",
3
+ "engine": "Isaac Sim",
4
+ "task": "Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.",
5
+ "inputs": [
6
+ "video.mp4",
7
+ "scale_anchor.json",
8
+ "anchor_reference.png",
9
+ "anchor_marked.png",
10
+ "self_check.py"
11
+ ],
12
+ "prompt": "prompt.txt",
13
+ "units": "metres, kilograms, seconds",
14
+ "up_axis": "+Z",
15
+ "entrypoint": "scene.usda",
16
+ "budget": {
17
+ "wall_clock_seconds": 1800,
18
+ "scope": "per_scene",
19
+ "starts_at": "agent_input_access",
20
+ "includes_all_tool_and_validation_time": true,
21
+ "deadline_action": "stop and submit latest valid loadable output",
22
+ "enforcement": "Docker generation runner terminates the container at the wall-clock deadline"
23
+ },
24
+ "requirement": "Sim-ready: ready for interactive robot learning, with no manual repair.",
25
+ "sim_ready": {
26
+ "loads_headless_offline": true,
27
+ "resets_to_rest": true,
28
+ "collision_matches_visible_geometry": true,
29
+ "rigid_things": "rigid bodies; free objects movable, articulated parts on limited joints",
30
+ "soft_things": [
31
+ "cloth",
32
+ "bedding",
33
+ "curtains",
34
+ "rugs",
35
+ "throws",
36
+ "cushions",
37
+ "pillows"
38
+ ],
39
+ "soft_mechanism": "native deformable",
40
+ "stable_under_sustained_interaction": true
41
+ },
42
+ "visual_objective": "Maximize visual quality, realism and fidelity to the source video, keeping every sim-ready requirement.",
43
+ "refinement": {
44
+ "minimum_completed_rounds": 3,
45
+ "minimum_reference_views": 6,
46
+ "minimum_inspected_video_frames": 12,
47
+ "final_validation_seconds": 90,
48
+ "outside_data_allowed": false
49
+ },
50
+ "label": "free text saying plainly what the object is; one entry per physical instance",
51
+ "interaction_points": [
52
+ "handle",
53
+ "grasp",
54
+ "spout",
55
+ "material_corners"
56
+ ],
57
+ "construction_rules": {
58
+ "max_initial_penetration_m": 0.005,
59
+ "max_penetration_duration_s": 0.02,
60
+ "max_passive_speed_mps": 5.0,
61
+ "articulated_parts_clear_of_carcass": true,
62
+ "swing_paths_clear": true,
63
+ "timestep": "agent's choice"
64
+ },
65
+ "self_check": "/isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check"
66
+ }
scenes/isaac/47331202/20261114T065042Z/input/prompt.txt ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ About the simulator. NVIDIA Isaac Sim is a robotics simulator built on Omniverse. You describe a scene
2
+ as an OpenUSD stage: prims carrying physics schemas for rigid bodies, colliders, joints and articulations, and
3
+ deformable bodies and surfaces. NVIDIA PhysX simulates it, here with GPU dynamics, and RTX path tracing renders it.
4
+ Isaac Sim is widely used for robot learning and sim-to-real work.
5
+
6
+ Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.
7
+
8
+ THE RESULT MUST BE SIM-READY, READY FOR INTERACTIVE ROBOT LEARNING. It is an environment a robot learns
9
+ in, not a picture of a room. A training pipeline must be able to load it headless, reset it, step it for
10
+ a long time and let robots act on it, with no manual repair. That means:
11
+ - It loads with no network access or external files, and resets to the same state at rest.
12
+ - Everything a robot could touch has collision geometry matching what is visible. An object with no
13
+ collision does not exist for a robot, however good it looks.
14
+ - Rigid things are rigid bodies with plausible mass; free objects can be pushed, and doors, drawers and lids
15
+ move on joints with limits.
16
+ - Soft things are native deformables: cloth, bedding, curtains, rugs, throws, cushions and pillows.
17
+ A rigid body that merely looks soft is wrong.
18
+ - It stays stable under sustained interaction, across many steps without a reset, and steps fast
19
+ enough to train on.
20
+
21
+ Budget: 30 minutes (1,800 seconds) of wall-clock time per scene, starting when
22
+ you receive access to the input files. All inspection, coding, asset creation,
23
+ tool calls, rendering, debugging, and validation must fit within this budget.
24
+ Stop work at the deadline and submit the latest valid, loadable result.
25
+ Do not stop after producing the first runnable scene. Use the available time
26
+ for repeated self-critique and refinement; reserve the final 90 seconds for
27
+ validation and submission.
28
+
29
+ Inputs (in /work/input):
30
+ - video.mp4: the complete source video.
31
+ - scale_anchor.json: dimensions and identity of one object in the room, the scale anchor.
32
+ - anchor_reference.png and anchor_marked.png: images identifying that object.
33
+ The marked rectangle is a projected box envelope, not a segmentation mask.
34
+ - generation_spec.json: this contract in machine-readable form.
35
+ - self_check.py: the public self-check described below.
36
+
37
+ Maximize visual realism and fidelity to the source video. Reproduce the room's
38
+ geometry, proportions, native objects, materials, textures, colors, lighting,
39
+ and visible details as faithfully as possible within the time budget. Use the
40
+ scale anchor to establish metric scale; infer other dimensions and layout from
41
+ the video. The anchor's known real dimensions let you work out every other
42
+ object's size and the distances between objects, for example by measuring them
43
+ against the anchor in frames where both are visible, or by scaling a
44
+ reconstruction so that the anchor has its true size. Place the anchor object
45
+ where it is in the video; the room is not aligned to it alone.
46
+
47
+ Visual objective: MAXIMIZE THE VISUAL QUALITY of the reconstructed room, without giving up any
48
+ sim-ready requirement: a scene that looks right but cannot be simulated has failed.
49
+ Use source-aligned renders to improve what a viewer actually sees: recognizable
50
+ furniture silhouettes, correct proportions and placement, detailed surfaces,
51
+ plausible texture scale, window appearance, soft lighting and small distinctive
52
+ objects. Avoid a room made only of plain boxes when the video shows richer
53
+ shapes and materials. Create local textures or meshes when they improve fidelity;
54
+ do not use a single photograph or video plane as a substitute for 3D geometry.
55
+ Check the entire camera path, including corners and reverse views, rather than
56
+ optimizing one attractive render.
57
+
58
+ Required self-critique and refinement:
59
+ 1. Inspect at least 12 frames distributed across the video. Choose at least
60
+ six representative reference views covering the room and important objects.
61
+ Record their timestamps. Estimate corresponding cameras using ONLY the
62
+ public video and scale anchor, and reuse those cameras across iterations.
63
+ 2. Build an early loadable baseline and render those views. Do not spend the
64
+ whole budget on a first draft or a single attractive camera angle.
65
+ 3. Complete at least THREE render-critique-refine-verify rounds. In each round,
66
+ open side-by-side source/render contact sheets with the image-capable Read
67
+ tool. Identify concrete visible discrepancies, implement the highest-impact
68
+ fixes, then re-render the SAME views and check for regressions. Writing a
69
+ critique without changing and re-rendering the scene is not a completed round.
70
+ 4. Round 1 prioritizes room shape, scale, furniture placement, object dimensions
71
+ and silhouettes. Round 2 prioritizes materials, textures, lighting, colors,
72
+ windows and salient details. Round 3 revisits visual mismatches and verifies
73
+ collision support, movable parts and soft items, and runs the self-check.
74
+ 5. Continue additional rounds while time permits. Do not wait until the deadline
75
+ to save a valid scene. Report the true number of completed rounds.
76
+ 6. After each round, save a valid checkpoint and record the source timestamps,
77
+ render paths, specific discrepancies, actual edits, before/after observations,
78
+ and remaining issues in refinement_log.json. Treat self-assessments as
79
+ diagnostics.
80
+ 7. Keep the best validated complete scene, and restore it if a later change
81
+ regresses. The final output must describe that selected result.
82
+
83
+ You may use any tools, including web search, and the software installed here,
84
+ including Blender (`blender`). Download nothing from the internet (no files,
85
+ models, assets, datasets or packages), and do not look up this room, its video
86
+ or the dataset it comes from. Use no annotations, scans or other data about this
87
+ room beyond your inputs, no previous generated outputs and no other session's
88
+ reconstruction. Do not read authentication files. Do not add a robot.
89
+
90
+ The result must be a genuinely interactive 3D scene:
91
+ - Preserve native object identities, count, dimensions, and spatial relationships.
92
+ - Provide collision geometry consistent with visible surfaces and free spaces.
93
+ - Separate movable objects and articulated parts, with appropriate joints and limits.
94
+ - Build every visible soft item with the simulator's native deformable mechanism, as one object
95
+ entry each, with a label that says plainly what it is, for example "curtain".
96
+ - Preserve openings, container interiors, support surfaces, and access needed
97
+ for interactions. Do not replace the room's own objects with generic stand-ins.
98
+ - Assign physically plausible properties and document estimates for unknowns;
99
+ the video is not measured mass, friction or stiffness.
100
+
101
+ Construction rules (the self-check reports them):
102
+ - The scene must start at rest without interpenetration: no contact deeper than
103
+ 5 mm between any two bodies or between a body and fixed geometry.
104
+ - Articulated parts (drawers, doors, lids) must not intersect their own carcass in
105
+ any position of their range: hollow the carcass or exclude the pair from collision.
106
+ - Door and drawer swing paths must be clear of other furniture.
107
+ - The timestep is your choice.
108
+
109
+ Object labels: each objects.json entry has a "label" that says plainly what the object is
110
+ (for example "double bed", "floor lamp"). Use one entry per physical instance.
111
+
112
+ Interaction points: for articulated furniture give a "handle" point on each moving
113
+ part; for movable graspable objects a "grasp" point; for pourable containers a
114
+ "spout" point; for a visible bed cover, four "material_corners" in boundary order.
115
+
116
+ Self-check: run it on your output after each round and before submitting. It loads
117
+ only your output, simulates 3 s from the authored state and reports contacts deeper
118
+ than 5 mm (with the two bodies involved), the time above 5 mm, peak body speed,
119
+ non-finite state and entries without a label. The scene must stay within these
120
+ limits: at most 0.02 s above 5 mm, peak speed at most 5 m/s, finite state. Fix what it reports; do not edit it.
121
+ /isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check
122
+
123
+ Output format (Isaac Sim, in /work/output):
124
+ - scene.usda (or scene.usd): self-contained USD stage, Z-up, metres, explicit default
125
+ prim, locally referenced assets and textures, and an authored PhysX scene.
126
+ - build_scene.py: reproducible builder using the installed Isaac Sim / pxr APIs.
127
+ - objects.json: {"objects": [...]}, one entry per physical object instance:
128
+ "id": a persistent unique string (keep the scale anchor ID),
129
+ "label": what the object is, in plain words,
130
+ "prim_path": the object's root prim (no single mesh containing all furniture),
131
+ "joint_paths": articulated joint prims, if present,
132
+ "interaction_points": interaction Xform prim paths, named handle, grasp or spout
133
+ (child Xforms of the moving part or object),
134
+ "material_corners": four mesh vertex indices for a deformable bed cover.
135
+ - render_preview.py, at least six views rendered with RTX PathTracing, preview.mp4
136
+ (camera tour) and interaction.mp4 (native physics interaction with an original object).
137
+ - README.md and refinement_log.json, as above.
138
+ Runtime: /isaac-sim/python.sh YOUR_SCRIPT.py runs the bundled Python with Isaac Sim,
139
+ USD, NumPy and Pillow; /usr/local/bin/ffmpeg and blender are available. Use headless=True,
140
+ multi_gpu=False, active_gpu=0, physics_gpu=0. sdk_smoke.py is an API and rendering
141
+ example. Native RTX capture can need several app updates before RGB has shape
142
+ (height, width, 4). Only input and output are host-mounted. Do not access
143
+ authentication.
scenes/isaac/47331202/20261114T065042Z/input/scale_anchor.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "object_id": "scale_anchor",
3
+ "description": "The table outlined in red in the marked image.",
4
+ "reference_image": "anchor_reference.png",
5
+ "marked_image": "anchor_marked.png",
6
+ "region_xyxy_pixels": [
7
+ 50.18202989863427,
8
+ 9.840496777454728,
9
+ 212.63893423128675,
10
+ 185.5016786254594
11
+ ],
12
+ "region_type": "projected_3d_box_envelope_not_segmentation",
13
+ "size_local_xyz_m": [
14
+ 0.803188210828418,
15
+ 1.314670830811198,
16
+ 0.7819756633824008
17
+ ],
18
+ "axis_description": [
19
+ "horizontal depth",
20
+ "horizontal width",
21
+ "vertical height"
22
+ ],
23
+ "measurement_type": "full extents of an oriented box around the object",
24
+ "uncertainty_m": null,
25
+ "uncertainty_note": "No uncertainty is supplied for these dimensions."
26
+ }
scenes/isaac/47331202/20261114T065042Z/input/sdk_smoke.py ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Verify Isaac Sim headless RTX rendering and native rigid-body stepping."""
2
+ from isaacsim import SimulationApp
3
+ app = SimulationApp({'headless':True,'renderer':'PathTracing','samples_per_pixel_per_frame':32,
4
+ 'width':512,'height':384,'multi_gpu':False,'active_gpu':0,'physics_gpu':0,
5
+ 'enable_crashreporter':False,'limit_cpu_threads':8})
6
+ import json
7
+ from pathlib import Path
8
+ import numpy as np
9
+ from PIL import Image
10
+ import omni.usd
11
+ import omni.replicator.core as rep
12
+ rep.orchestrator.set_capture_on_play(False)
13
+ from pxr import UsdGeom,UsdLux,UsdShade,UsdPhysics,Gf,Sdf
14
+ stage=omni.usd.get_context().get_stage()
15
+ UsdGeom.SetStageUpAxis(stage,UsdGeom.Tokens.z);UsdGeom.SetStageMetersPerUnit(stage,1.)
16
+ UsdGeom.Xform.Define(stage,'/World');stage.SetDefaultPrim(stage.GetPrimAtPath('/World'))
17
+ def box(path,pos,scale,color,dynamic=False):
18
+ prim=UsdGeom.Cube.Define(stage,path);prim.CreateSizeAttr(1.)
19
+ xf=UsdGeom.Xformable(prim);xf.AddTranslateOp().Set(Gf.Vec3d(*pos));xf.AddScaleOp().Set(Gf.Vec3f(*scale))
20
+ UsdPhysics.CollisionAPI.Apply(prim.GetPrim())
21
+ if dynamic:
22
+ UsdPhysics.RigidBodyAPI.Apply(prim.GetPrim());UsdPhysics.MassAPI.Apply(prim.GetPrim()).CreateMassAttr(1.)
23
+ mat=UsdShade.Material.Define(stage,path+'_Material');shader=UsdShade.Shader.Define(stage,path+'_Material/Shader');shader.CreateIdAttr('UsdPreviewSurface')
24
+ shader.CreateInput('diffuseColor',Sdf.ValueTypeNames.Color3f).Set(Gf.Vec3f(*color));shader.CreateInput('roughness',Sdf.ValueTypeNames.Float).Set(.28)
25
+ mat.CreateSurfaceOutput().ConnectToSource(shader.ConnectableAPI(),'surface');UsdShade.MaterialBindingAPI.Apply(prim.GetPrim()).Bind(mat)
26
+ return prim
27
+ box('/World/Floor',(0,0,-.05),(5,5,.1),(.6,.6,.6));box('/World/Cube',(0,0,1.),(.3,.3,.3),(.08,.3,.8),True)
28
+ light=UsdLux.DomeLight.Define(stage,'/World/Dome');light.CreateIntensityAttr(500.)
29
+ light=UsdLux.RectLight.Define(stage,'/World/Key');light.CreateIntensityAttr(2500.);light.CreateWidthAttr(3.);light.CreateHeightAttr(3.);UsdGeom.Xformable(light).AddTranslateOp().Set(Gf.Vec3d(0,0,3))
30
+ scene=UsdPhysics.Scene.Define(stage,'/World/PhysicsScene');scene.CreateGravityDirectionAttr(Gf.Vec3f(0,0,-1));scene.CreateGravityMagnitudeAttr(9.81)
31
+ camera=rep.create.camera(position=(2.4,-3,2),look_at=(0,0,.3));product=rep.create.render_product(camera,(512,384));rgb=rep.AnnotatorRegistry.get_annotator('rgb');rgb.attach([product])
32
+ from isaacsim.core.api import World
33
+ world=World(physics_dt=1/60,rendering_dt=1/60,stage_units_in_meters=1.)
34
+ world.reset()
35
+ for _ in range(30):world.step(render=True)
36
+ for attempt in range(30):
37
+ rep.orchestrator.step(rt_subframes=8)
38
+ app.update()
39
+ image=np.asarray(rgb.get_data())
40
+ print('CAPTURE_ATTEMPT',attempt,'SHAPE',image.shape,flush=True)
41
+ if image.ndim==3 and image.shape[0]>0:break
42
+ print('RGB_SHAPE',image.shape,flush=True)
43
+ assert image.ndim==3,'RTX annotator did not return a frame'
44
+ position=UsdGeom.Xformable(stage.GetPrimAtPath('/World/Cube')).ComputeLocalToWorldTransform(0).ExtractTranslation()
45
+ print('CUBE_POSITION',list(position),flush=True)
46
+ out=Path('/work/output');out.mkdir(exist_ok=True)
47
+ Image.fromarray(image[:,:,:3]).save(out/'smoke.png');stage.GetRootLayer().Export(str(out/'sdk_example.usda'))
48
+ (out/'smoke.json').write_text(json.dumps({'renderer':'Isaac Sim RTX PathTracing','shape':list(image.shape),'cube_final_position':list(position),'initial_height_m':1.,'success':True}))
49
+ app.close()
scenes/isaac/47331202/20261114T065042Z/input/self_check.py ADDED
@@ -0,0 +1,228 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Passive physics of a USD scene in Isaac Sim, against the scene requirements (no penetration beyond 5 mm for more than
2
+ 0.02 s, no body faster than 5 m/s, finite state).
3
+
4
+ Runs inside the Isaac Sim container. Reads --scene (default /work/scene, read-only) and the optional
5
+ --out/request.json (object ids to drop); writes --out/physics.json (default /work/physics). Without a request it is
6
+ the self-check: no drops, plus a check that every object has a non-empty label.
7
+
8
+ - Speed: a 3 s passive rollout with no rendering and no state reads, timed around world.step only
9
+ (simulated seconds / wall seconds).
10
+ - Passive rollout: 3 s from the authored state. Penetration comes from the PhysX tensor contact view
11
+ (GPU contact reports return nothing under GPU dynamics); speed and poses from the RigidPrim view.
12
+ - Drops: the requested objects with free (unjointed) dynamic bodies, raised 0.05 m
13
+ and released for 2 s, each from a fresh reset.
14
+ The limits are the same as for MuJoCo scenes.
15
+ """
16
+ from isaacsim import SimulationApp
17
+ app=SimulationApp({'headless':True,'multi_gpu':False,'active_gpu':0,'physics_gpu':0,
18
+ 'enable_crashreporter':False,'limit_cpu_threads':8})
19
+ import argparse
20
+ import json
21
+ from pathlib import Path
22
+ import time
23
+ import carb
24
+ import numpy as np
25
+ import omni.usd
26
+ from pxr import UsdPhysics,PhysxSchema
27
+
28
+ LIMIT_M,LIMIT_S,SPEED=.005,.02,5.
29
+ parser=argparse.ArgumentParser();parser.add_argument('--scene',default='/work/scene');parser.add_argument('--out',default='/work/physics')
30
+ args,_=parser.parse_known_args()
31
+ out=Path(args.out);out.mkdir(parents=True,exist_ok=True)
32
+ request=json.loads((out/'request.json').read_text()) if (out/'request.json').is_file() else {'object_ids':[]}
33
+ result={'status':'failed','engine':'Isaac Sim PhysX (GPU dynamics)','no_vlm':True,'provisional':True,
34
+ 'thresholds':{'penetration_m':LIMIT_M,'penetration_duration_s':LIMIT_S,'linear_speed_mps':SPEED},
35
+ 'thresholds_calibrated_across_engines':False}
36
+
37
+
38
+ def contact_filters(stage,paths):
39
+ """GPU contact views filter on rigid bodies or standalone colliders only, so a collider inside a
40
+ rigid body is replaced by that body (unsupported filters are dropped with only a warning)."""
41
+ out=[]
42
+ for path in paths:
43
+ if any('deformable' in str(a).lower() for a in stage.GetPrimAtPath(path).GetAppliedSchemas()):
44
+ continue # deformable meshes cannot be contact filters; their contacts are not measured
45
+ prim=stage.GetPrimAtPath(path);owner=None
46
+ while prim and not prim.IsPseudoRoot():
47
+ if prim.HasAPI(UsdPhysics.RigidBodyAPI):owner=str(prim.GetPath())
48
+ prim=prim.GetParent()
49
+ target=owner or path
50
+ if target not in out:out.append(target)
51
+ return out
52
+
53
+
54
+ def array(value):
55
+ return np.asarray(value.cpu() if hasattr(value,'cpu') else value,dtype=float)
56
+
57
+
58
+ try:
59
+ root=Path(args.scene)
60
+ scene=next(root/n for n in ['scene.usda','scene.usd'] if (root/n).is_file())
61
+ omni.usd.get_context().open_stage(str(scene));stage=omni.usd.get_context().get_stage()
62
+ for _ in range(8):app.update()
63
+
64
+ def dynamic(prim):
65
+ api=UsdPhysics.RigidBodyAPI(prim)
66
+ if api.GetRigidBodyEnabledAttr().Get() not in (None,True) or api.GetKinematicEnabledAttr().Get():return False
67
+ parent=prim.GetParent()
68
+ while parent and not parent.IsPseudoRoot():
69
+ if parent.HasAPI(UsdPhysics.RigidBodyAPI):return False
70
+ parent=parent.GetParent()
71
+ return True
72
+ rigid=[p for p in stage.Traverse() if p.HasAPI(UsdPhysics.RigidBodyAPI)]
73
+ bodies=[p for p in rigid if dynamic(p)]
74
+ paths=[str(p.GetPath()) for p in bodies]
75
+ colliders=[str(p.GetPath()) for p in stage.Traverse() if p.HasAPI(UsdPhysics.CollisionAPI)]
76
+ jointed=set()
77
+ for prim in stage.Traverse():
78
+ if prim.IsA(UsdPhysics.Joint):
79
+ joint=UsdPhysics.Joint(prim)
80
+ for rel in (joint.GetBody0Rel(),joint.GetBody1Rel()):
81
+ jointed.update(str(t) for t in rel.GetTargets())
82
+ for prim in bodies:PhysxSchema.PhysxContactReportAPI.Apply(prim).CreateThresholdAttr().Set(0.)
83
+ scenes=[p for p in stage.Traverse() if p.IsA(UsdPhysics.Scene)]
84
+ authored=PhysxSchema.PhysxSceneAPI(scenes[0]).GetTimeStepsPerSecondAttr().Get() if scenes else None
85
+ rate=float(authored or 60)
86
+ result.update(dynamic_bodies=len(paths),colliders=len(colliders),
87
+ excluded_bodies=[str(p.GetPath()) for p in rigid if not dynamic(p)],
88
+ timestep_s=1/rate,timestep_source='authored PhysxSceneAPI timeStepsPerSecond' if authored else 'default 60 Hz')
89
+ if not paths:raise ValueError('No dynamic rigid bodies to simulate')
90
+ filters=contact_filters(stage,colliders)
91
+ result['contact_filters']=len(filters)
92
+
93
+ carb.settings.get_settings().set_bool('/physics/disableContactProcessing',False)
94
+ from isaacsim.core.api import World
95
+ from isaacsim.core.prims import RigidPrim
96
+ from isaacsim.core.simulation_manager import SimulationManager
97
+ world=World(physics_dt=1/rate,rendering_dt=1/rate,stage_units_in_meters=1.,
98
+ physics_prim_path=str(scenes[0].GetPath()) if scenes else '/World/AddedPhysicsScene',
99
+ set_defaults=False,backend='torch',device='cuda:0')
100
+ context=world.get_physics_context()
101
+ context.enable_gpu_dynamics(True);context.set_broadphase_type('GPU')
102
+ view=RigidPrim(paths,name='scene_bodies');world.scene.add(view)
103
+ steps=lambda seconds:int(np.ceil(seconds*rate))
104
+
105
+ def measure_rtf(dt,repeats=3):
106
+ """Simulated seconds / wall seconds of world.step(render=False) for an unread passive rollout from the authored
107
+ state, median over `repeats` after 10 warm-up steps."""
108
+ world.set_simulation_dt(physics_dt=dt,rendering_dt=dt)
109
+ world.reset()
110
+ for _ in range(10):world.step(render=False)
111
+ count=int(np.ceil(3/dt));samples=[];finite=True
112
+ for _ in range(repeats):
113
+ world.reset();start=time.perf_counter()
114
+ for _ in range(count):world.step(render=False)
115
+ samples.append(count*dt/(time.perf_counter()-start))
116
+ positions,_=view.get_world_poses()
117
+ finite&=bool(np.isfinite(array(positions)).all() and np.isfinite(array(view.get_velocities())).all())
118
+ return {'rtf':float(np.median(samples)),'rtf_samples':samples,'timestep_s':dt,'simulated_seconds':count*dt,
119
+ 'steps':count,'repeats':repeats,'finite':finite,
120
+ 'definition':'simulated seconds / wall seconds of world.step(render=False) only, median of repeats'}
121
+
122
+ result['realtime']={**measure_rtf(1/rate),'scope':'3 s passive rollout from the authored state, authored timestep'}
123
+ print('PHYSICS_STAGE realtime',result['realtime']['rtf'],flush=True)
124
+
125
+ def reset():
126
+ world.reset()
127
+ contacts=SimulationManager.get_physics_sim_view().create_rigid_contact_view(
128
+ paths,filter_patterns=[filters]*len(paths),max_contact_data_count=200000)
129
+ if contacts is None:raise RuntimeError('PhysX rigid contact view could not be created')
130
+ return contacts
131
+
132
+ def rollout(seconds,contacts,motion=None):
133
+ """One rollout; `motion` names the file the logged poses are written to."""
134
+ report={'finite':True,'peak_penetration_m':0.,'penetration_time_s':0.,'peak_linear_speed_mps':0.}
135
+ every=max(1,round(rate/10));frames=[]
136
+ for step in range(steps(seconds)):
137
+ world.step(render=False)
138
+ positions,orientations=view.get_world_poses();velocities=array(view.get_velocities())
139
+ if step%every==0:frames.append(((step+1)/rate,array(positions),array(orientations)))
140
+ report['finite']&=bool(np.isfinite(array(positions)).all() and np.isfinite(velocities).all())
141
+ _,_,_,separation,counts,starts=contacts.get_contact_data(dt=1/rate)
142
+ total=int(array(counts).sum())
143
+ if 'initial_pairs' not in report:
144
+ # Pairs deeper than the limit after the first step, named for the self-check.
145
+ pairs=[];sep=array(separation);cnt=array(counts).astype(int);beg=array(starts).astype(int)
146
+ for i,j in zip(*np.nonzero(cnt)):
147
+ depth=-float(sep[beg[i,j]:beg[i,j]+cnt[i,j]].min())
148
+ if depth>LIMIT_M:pairs.append({'depth_m':depth,'bodies':[paths[i],filters[j]]})
149
+ report['initial_pairs']=sorted(pairs,key=lambda p:-p['depth_m'])[:20]
150
+ deepest=max(0.,-float(array(separation)[:total].min())) if total else 0.
151
+ report['peak_penetration_m']=max(report['peak_penetration_m'],deepest)
152
+ if deepest>LIMIT_M:report['penetration_time_s']+=1/rate
153
+ report['peak_linear_speed_mps']=max(report['peak_linear_speed_mps'],float(np.linalg.norm(velocities[:,:3],axis=1).max()))
154
+ if not report['finite']:break
155
+ report['valid']=bool(report['finite'] and report['penetration_time_s']<=LIMIT_S and report['peak_linear_speed_mps']<=SPEED)
156
+ report['thresholds']=result['thresholds']
157
+ if motion and request.get('record'):
158
+ (out/'motion').mkdir(exist_ok=True)
159
+ np.savez_compressed(out/'motion'/motion,paths=np.array(paths),t=np.array([f[0] for f in frames]),
160
+ positions=np.stack([f[1] for f in frames]),orientations=np.stack([f[2] for f in frames]))
161
+ report['motion']=f'motion/{motion}.npz'
162
+ return report
163
+
164
+ contacts=reset()
165
+ start_positions=array(view.get_world_poses()[0])
166
+ masses=array(view.get_masses())
167
+ direction,magnitude=context.get_gravity()
168
+ gravity=np.asarray(direction,dtype=float)*float(magnitude)
169
+ passive=rollout(3,contacts,'passive')
170
+ print('PHYSICS_STAGE passive',passive,flush=True)
171
+ end_positions=array(view.get_world_poses()[0])
172
+ moved=np.linalg.norm(end_positions-start_positions,axis=1)
173
+
174
+ manifest=json.loads((root/'objects.json').read_text())
175
+ manifest=manifest['objects'] if isinstance(manifest,dict) else manifest
176
+ result['manifest_errors']=[f"{o.get('id')}: label is missing or empty" for o in manifest if not str(o.get('label') or '').strip()]
177
+ result['manifest_errors']+=[f"{o.get('id')}: prim_path does not resolve" for o in manifest if not stage.GetPrimAtPath(o.get('prim_path','') or '/')]
178
+ def owned(obj):
179
+ prefix=str(obj.get('prim_path','')).rstrip('/')
180
+ return [i for i,p in enumerate(paths) if prefix and (p==prefix or p.startswith(prefix+'/'))]
181
+ drift=[{'id':o['id'],'dynamic_bodies':len(owned(o)),'center_drift_m':float(moved[owned(o)].max()),'mass_kg':float(masses[owned(o)].sum())}
182
+ for o in manifest if owned(o)]
183
+
184
+ requested=set(request['object_ids'])
185
+ eligible=[o for o in manifest if o['id'] in requested]
186
+ drops=[]
187
+ for obj in eligible:
188
+ free=[i for i in owned(obj) if paths[i] not in jointed]
189
+ if not free:continue
190
+ contacts=reset()
191
+ positions,orientations=view.get_world_poses()
192
+ raised=positions.clone() if hasattr(positions,'clone') else positions.copy()
193
+ raised[free,2]+=.05
194
+ view.set_world_poses(positions=raised,orientations=orientations)
195
+ applied=float((array(view.get_world_poses()[0])[free,2]-array(positions)[free,2]).min())
196
+ drops.append({'id':obj['id'],'input':'raise free bodies 0.05 m and release','bodies':[paths[i] for i in free],
197
+ 'lift_applied_m':applied,**rollout(2,contacts,f'drop_{len(drops)}')})
198
+ print('PHYSICS_STAGE drop',obj['id'],applied,drops[-1]['peak_linear_speed_mps'],flush=True)
199
+
200
+ # The common 2 ms step last, so the rollouts above run at the authored step.
201
+ result['realtime_common_step']={**measure_rtf(.002),'scope':'3 s passive rollout from the authored state at the common 2 ms step'}
202
+ print('PHYSICS_STAGE realtime_common_step',result['realtime_common_step']['rtf'],flush=True)
203
+ gravity_ok=bool(np.allclose(gravity,[0,0,-9.81],atol=.1))
204
+ mass_ok=bool(np.isfinite(masses).all() and (masses>=0).all())
205
+ result.update(status='completed',passive=passive,gravity_m_s2=gravity.tolist(),gravity_valid=gravity_ok,
206
+ mass_valid=mass_ok,center_drift=drift,drop_tests=drops,eligible_native_objects=len(eligible),
207
+ largest_body_displacements=[{'prim':paths[i],'displacement_m':float(moved[i])} for i in np.argsort(-moved)[:5]],
208
+ valid=bool(passive['valid'] and gravity_ok and mass_ok),
209
+ drop_scope='the requested objects with free (unjointed) dynamic bodies, in manifest order',
210
+ contact_scope=('contacts between each dynamic body and every collider (static, kinematic and dynamic; colliders '
211
+ 'inside a rigid body are filtered by that body, as GPU contact views require); '
212
+ 'static-static pairs are not simulated, as in MuJoCo world-body geoms'),
213
+ limitations=['the limits are the MuJoCo ones, not calibrated for PhysX',
214
+ 'kinematic, disabled and nested rigid bodies are not traced (see excluded_bodies)',
215
+ 'drift is per dynamic body under the object prim, not the object bounding box',
216
+ 'contacts with deformable bodies are not measured (GPU contact views cannot filter them)'])
217
+ except Exception as error:
218
+ result['error']=f'{type(error).__name__}: {error}'
219
+ finally:
220
+ (out/'physics.json').write_text(json.dumps(result,indent=2)+'\n')
221
+ if 'passive' in result:
222
+ passive=result['passive']
223
+ print('PASSIVE CHECK','PASSED' if passive['valid'] and not result.get('manifest_errors') else 'FAILED',
224
+ f"- {len(passive.get('initial_pairs',[]))} contact pairs > 5 mm, {passive['penetration_time_s']:.3f} s above 5 mm,"
225
+ f" peak speed {passive['peak_linear_speed_mps']:.2f} m/s, {len(result.get('manifest_errors',[]))} manifest errors;"
226
+ f" details in {out/'physics.json'}",flush=True)
227
+ else:print('CHECK FAILED:',result.get('error'),flush=True)
228
+ app.close()
scenes/isaac/47331202/20261114T065042Z/input/video.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:066fcbfb5a6f9ffecd6af6845e654de29195adc2c3d233a9092a11c8b81c0e2a
3
+ size 11908131
scenes/isaac/47331202/20261114T065042Z/run.json ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "track": "scenes",
3
+ "engine": "isaac",
4
+ "item": "47331202",
5
+ "run_id": "20261114T065042Z",
6
+ "campaign": "jiarui-2-codebuddy",
7
+ "cli": "codebuddy",
8
+ "model": "Hunyuan 4",
9
+ "model_id": "hy4-preview",
10
+ "model_family": "hy4-preview",
11
+ "model_reasoning_effort": "default",
12
+ "budget_seconds": 1800,
13
+ "told_budget_seconds": 1800,
14
+ "stage": 1,
15
+ "continued_from": null,
16
+ "contract": {
17
+ "name": "scenes-isaac",
18
+ "sha256": "f1ab7ea88842502d37176b41d9f3bec64a5d9fac292ca3f16caf08da3b5db96d",
19
+ "track": "scenes",
20
+ "engine": "isaac",
21
+ "engine_version": "6.2.0",
22
+ "image": "simready-isaac-generation:latest",
23
+ "blender_version": "5.2.3",
24
+ "inputs": [
25
+ "video.mp4",
26
+ "scale_anchor.json",
27
+ "anchor_reference.png",
28
+ "anchor_marked.png",
29
+ "prompt.txt",
30
+ "generation_spec.json",
31
+ "self_check.py",
32
+ "sdk_smoke.py"
33
+ ],
34
+ "withheld": [
35
+ "groundtruth/ (objects, cameras, task catalogs, object dynamics, interaction test) and its source tag",
36
+ "provenance/ (source annotation, anchor selection)",
37
+ "the frozen protocol of the track and engine (data/benchmark/frozen/<track>/<engine>.json)",
38
+ "other runs"
39
+ ],
40
+ "outputs": [
41
+ "scene.usda",
42
+ "build_scene.py",
43
+ "objects.json",
44
+ "render_preview.py",
45
+ "preview.mp4",
46
+ "interaction.mp4",
47
+ "README.md",
48
+ "refinement_log.json"
49
+ ],
50
+ "self_check": "/isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check",
51
+ "budget": {
52
+ "told": "the first of the session's cumulative budgets, in wall-clock seconds from container start",
53
+ "resume": "each later budget resumes the same conversation for the rest of its time; each budget is its own run",
54
+ "enforcement": "the runner kills the container at the deadline; the output as it stands is the result"
55
+ },
56
+ "refinement": {
57
+ "minimum_completed_rounds": 3,
58
+ "minimum_reference_views": 6,
59
+ "minimum_inspected_video_frames": 12,
60
+ "final_validation_seconds": 90
61
+ },
62
+ "isolation": {
63
+ "container": "Docker: read-only input (and tools), writable output and scratch, temporary credentials; no repository, benchmark ground truth or evaluator files",
64
+ "network": "an internal network whose only exit is an HTTPS CONNECT proxy (runtime/egress_proxy.py)",
65
+ "egress_hosts": [
66
+ "api.anthropic.com",
67
+ "platform.claude.com",
68
+ "chatgpt.com",
69
+ "auth.openai.com",
70
+ "opencode.ai",
71
+ "mcp.exa.ai",
72
+ "search.parallel.ai"
73
+ ]
74
+ },
75
+ "protocol": {
76
+ "file": "data/benchmark/frozen/scenes/isaac.json",
77
+ "sha256": "df0fe5bb1c796474ad13cdba81b660a2946aa24297cb289f9581cfde7c035bb7"
78
+ }
79
+ },
80
+ "engine_version": "6.2.0",
81
+ "image": "simready-isaac-generation:latest",
82
+ "image_id": "sha256:8d9f9c934457aa3c8e91e8099e35db27b50db287d0eb76a647052c310e754d87",
83
+ "required_refinement_rounds": 3,
84
+ "gpu_uuid": "GPU-91246d27-fecc-875a-e18a-afb185bc6a18",
85
+ "egress_mode": "internal_network",
86
+ "input_snapshot": "input",
87
+ "tools": null,
88
+ "evaluator_protocol": "evaluator_protocol.json",
89
+ "output": "output",
90
+ "log": "generation.jsonl",
91
+ "status": "budget_expired",
92
+ "agent_cli": {
93
+ "cli": "codebuddy",
94
+ "name": "CodeBuddy Code",
95
+ "binary": "a pinned codebuddy-code package and its node, mounted read-only (SIMREADY_CODEBUDDY_PREFIX, default the host's codebuddy install)",
96
+ "login": "the CodeBuddy login file, given to the CLI as a snapshot without its refresh token (the session cannot rotate it)",
97
+ "billing": "CodeBuddy account (the CLI reports no price)",
98
+ "tools": "every tool but WebFetch and the server-side media generators ImageGen, ImageEdit, VideoGen (--disallowedTools); WebSearch runs on CodeBuddy's side",
99
+ "hosts": [
100
+ "[redacted-host]",
101
+ "wb.[redacted]buddy.com"
102
+ ],
103
+ "version": "2.182.0",
104
+ "model": "hy4-preview",
105
+ "model_family": "hy4-preview"
106
+ },
107
+ "started_at": "2026-11-14T06:50:43.909864+00:00",
108
+ "model_answered": true,
109
+ "exit_code": 137,
110
+ "credential_refresh_persistence": "snapshot_without_refresh_token",
111
+ "finished_at": "2026-11-14T07:20:46.272377+00:00",
112
+ "cli_session_id": "01a1d8d2-a69c-7b1c-b044-5f496bf53e5c",
113
+ "api_retries": 0,
114
+ "rate_limit_events": 0,
115
+ "stage_seconds": 1802.429,
116
+ "elapsed_seconds": 1802.429,
117
+ "cost": {
118
+ "status": "partial",
119
+ "source": "the CLI's running usage totals (usage_update events)",
120
+ "usd": null,
121
+ "usd_basis": "not reported (CodeBuddy account billing)",
122
+ "credits": 0,
123
+ "credit_status": "known",
124
+ "turns": null,
125
+ "api_seconds": null,
126
+ "tokens": {
127
+ "input": 65998,
128
+ "cache_read": 597248,
129
+ "cache_write": 0,
130
+ "output": 13008,
131
+ "reasoning": null,
132
+ "total": 676254
133
+ },
134
+ "wall_seconds": 1802.429,
135
+ "budget_seconds": 1800,
136
+ "stopped_at_budget": true
137
+ },
138
+ "scene_present": false
139
+ }
scenes/isaac/47331202/20261114T065042Z/submitted_prompt.txt ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ About the simulator. NVIDIA Isaac Sim is a robotics simulator built on Omniverse. You describe a scene
2
+ as an OpenUSD stage: prims carrying physics schemas for rigid bodies, colliders, joints and articulations, and
3
+ deformable bodies and surfaces. NVIDIA PhysX simulates it, here with GPU dynamics, and RTX path tracing renders it.
4
+ Isaac Sim is widely used for robot learning and sim-to-real work.
5
+
6
+ Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.
7
+
8
+ THE RESULT MUST BE SIM-READY, READY FOR INTERACTIVE ROBOT LEARNING. It is an environment a robot learns
9
+ in, not a picture of a room. A training pipeline must be able to load it headless, reset it, step it for
10
+ a long time and let robots act on it, with no manual repair. That means:
11
+ - It loads with no network access or external files, and resets to the same state at rest.
12
+ - Everything a robot could touch has collision geometry matching what is visible. An object with no
13
+ collision does not exist for a robot, however good it looks.
14
+ - Rigid things are rigid bodies with plausible mass; free objects can be pushed, and doors, drawers and lids
15
+ move on joints with limits.
16
+ - Soft things are native deformables: cloth, bedding, curtains, rugs, throws, cushions and pillows.
17
+ A rigid body that merely looks soft is wrong.
18
+ - It stays stable under sustained interaction, across many steps without a reset, and steps fast
19
+ enough to train on.
20
+
21
+ Budget: 30 minutes (1,800 seconds) of wall-clock time per scene, starting when
22
+ you receive access to the input files. All inspection, coding, asset creation,
23
+ tool calls, rendering, debugging, and validation must fit within this budget.
24
+ Stop work at the deadline and submit the latest valid, loadable result.
25
+ Do not stop after producing the first runnable scene. Use the available time
26
+ for repeated self-critique and refinement; reserve the final 90 seconds for
27
+ validation and submission.
28
+
29
+ Inputs (in /work/input):
30
+ - video.mp4: the complete source video.
31
+ - scale_anchor.json: dimensions and identity of one object in the room, the scale anchor.
32
+ - anchor_reference.png and anchor_marked.png: images identifying that object.
33
+ The marked rectangle is a projected box envelope, not a segmentation mask.
34
+ - generation_spec.json: this contract in machine-readable form.
35
+ - self_check.py: the public self-check described below.
36
+
37
+ Maximize visual realism and fidelity to the source video. Reproduce the room's
38
+ geometry, proportions, native objects, materials, textures, colors, lighting,
39
+ and visible details as faithfully as possible within the time budget. Use the
40
+ scale anchor to establish metric scale; infer other dimensions and layout from
41
+ the video. The anchor's known real dimensions let you work out every other
42
+ object's size and the distances between objects, for example by measuring them
43
+ against the anchor in frames where both are visible, or by scaling a
44
+ reconstruction so that the anchor has its true size. Place the anchor object
45
+ where it is in the video; the room is not aligned to it alone.
46
+
47
+ Visual objective: MAXIMIZE THE VISUAL QUALITY of the reconstructed room, without giving up any
48
+ sim-ready requirement: a scene that looks right but cannot be simulated has failed.
49
+ Use source-aligned renders to improve what a viewer actually sees: recognizable
50
+ furniture silhouettes, correct proportions and placement, detailed surfaces,
51
+ plausible texture scale, window appearance, soft lighting and small distinctive
52
+ objects. Avoid a room made only of plain boxes when the video shows richer
53
+ shapes and materials. Create local textures or meshes when they improve fidelity;
54
+ do not use a single photograph or video plane as a substitute for 3D geometry.
55
+ Check the entire camera path, including corners and reverse views, rather than
56
+ optimizing one attractive render.
57
+
58
+ Required self-critique and refinement:
59
+ 1. Inspect at least 12 frames distributed across the video. Choose at least
60
+ six representative reference views covering the room and important objects.
61
+ Record their timestamps. Estimate corresponding cameras using ONLY the
62
+ public video and scale anchor, and reuse those cameras across iterations.
63
+ 2. Build an early loadable baseline and render those views. Do not spend the
64
+ whole budget on a first draft or a single attractive camera angle.
65
+ 3. Complete at least THREE render-critique-refine-verify rounds. In each round,
66
+ open side-by-side source/render contact sheets with the image-capable Read
67
+ tool. Identify concrete visible discrepancies, implement the highest-impact
68
+ fixes, then re-render the SAME views and check for regressions. Writing a
69
+ critique without changing and re-rendering the scene is not a completed round.
70
+ 4. Round 1 prioritizes room shape, scale, furniture placement, object dimensions
71
+ and silhouettes. Round 2 prioritizes materials, textures, lighting, colors,
72
+ windows and salient details. Round 3 revisits visual mismatches and verifies
73
+ collision support, movable parts and soft items, and runs the self-check.
74
+ 5. Continue additional rounds while time permits. Do not wait until the deadline
75
+ to save a valid scene. Report the true number of completed rounds.
76
+ 6. After each round, save a valid checkpoint and record the source timestamps,
77
+ render paths, specific discrepancies, actual edits, before/after observations,
78
+ and remaining issues in refinement_log.json. Treat self-assessments as
79
+ diagnostics.
80
+ 7. Keep the best validated complete scene, and restore it if a later change
81
+ regresses. The final output must describe that selected result.
82
+
83
+ You may use any tools, including web search, and the software installed here,
84
+ including Blender (`blender`). Download nothing from the internet (no files,
85
+ models, assets, datasets or packages), and do not look up this room, its video
86
+ or the dataset it comes from. Use no annotations, scans or other data about this
87
+ room beyond your inputs, no previous generated outputs and no other session's
88
+ reconstruction. Do not read authentication files. Do not add a robot.
89
+
90
+ The result must be a genuinely interactive 3D scene:
91
+ - Preserve native object identities, count, dimensions, and spatial relationships.
92
+ - Provide collision geometry consistent with visible surfaces and free spaces.
93
+ - Separate movable objects and articulated parts, with appropriate joints and limits.
94
+ - Build every visible soft item with the simulator's native deformable mechanism, as one object
95
+ entry each, with a label that says plainly what it is, for example "curtain".
96
+ - Preserve openings, container interiors, support surfaces, and access needed
97
+ for interactions. Do not replace the room's own objects with generic stand-ins.
98
+ - Assign physically plausible properties and document estimates for unknowns;
99
+ the video is not measured mass, friction or stiffness.
100
+
101
+ Construction rules (the self-check reports them):
102
+ - The scene must start at rest without interpenetration: no contact deeper than
103
+ 5 mm between any two bodies or between a body and fixed geometry.
104
+ - Articulated parts (drawers, doors, lids) must not intersect their own carcass in
105
+ any position of their range: hollow the carcass or exclude the pair from collision.
106
+ - Door and drawer swing paths must be clear of other furniture.
107
+ - The timestep is your choice.
108
+
109
+ Object labels: each objects.json entry has a "label" that says plainly what the object is
110
+ (for example "double bed", "floor lamp"). Use one entry per physical instance.
111
+
112
+ Interaction points: for articulated furniture give a "handle" point on each moving
113
+ part; for movable graspable objects a "grasp" point; for pourable containers a
114
+ "spout" point; for a visible bed cover, four "material_corners" in boundary order.
115
+
116
+ Self-check: run it on your output after each round and before submitting. It loads
117
+ only your output, simulates 3 s from the authored state and reports contacts deeper
118
+ than 5 mm (with the two bodies involved), the time above 5 mm, peak body speed,
119
+ non-finite state and entries without a label. The scene must stay within these
120
+ limits: at most 0.02 s above 5 mm, peak speed at most 5 m/s, finite state. Fix what it reports; do not edit it.
121
+ /isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check
122
+
123
+ Output format (Isaac Sim, in /work/output):
124
+ - scene.usda (or scene.usd): self-contained USD stage, Z-up, metres, explicit default
125
+ prim, locally referenced assets and textures, and an authored PhysX scene.
126
+ - build_scene.py: reproducible builder using the installed Isaac Sim / pxr APIs.
127
+ - objects.json: {"objects": [...]}, one entry per physical object instance:
128
+ "id": a persistent unique string (keep the scale anchor ID),
129
+ "label": what the object is, in plain words,
130
+ "prim_path": the object's root prim (no single mesh containing all furniture),
131
+ "joint_paths": articulated joint prims, if present,
132
+ "interaction_points": interaction Xform prim paths, named handle, grasp or spout
133
+ (child Xforms of the moving part or object),
134
+ "material_corners": four mesh vertex indices for a deformable bed cover.
135
+ - render_preview.py, at least six views rendered with RTX PathTracing, preview.mp4
136
+ (camera tour) and interaction.mp4 (native physics interaction with an original object).
137
+ - README.md and refinement_log.json, as above.
138
+ Runtime: /isaac-sim/python.sh YOUR_SCRIPT.py runs the bundled Python with Isaac Sim,
139
+ USD, NumPy and Pillow; /usr/local/bin/ffmpeg and blender are available. Use headless=True,
140
+ multi_gpu=False, active_gpu=0, physics_gpu=0. sdk_smoke.py is an API and rendering
141
+ example. Native RTX capture can need several app updates before RGB has shape
142
+ (height, width, 4). Only input and output are host-mounted. Do not access
143
+ authentication.
scenes/isaac/47331202/20261114T083400Z/evaluator_protocol.json ADDED
The diff for this file is too large to render. See raw diff
 
scenes/isaac/47331202/20261114T083400Z/generation.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
scenes/isaac/47331202/20261114T083400Z/input/anchor_marked.png ADDED

Git LFS Details

  • SHA256: f4710332f02e78fd95d56162c4edb4dc504ae29bb4dc9bce768e4cfc177b446c
  • Pointer size: 130 Bytes
  • Size of remote file: 73.1 kB
scenes/isaac/47331202/20261114T083400Z/input/anchor_reference.png ADDED

Git LFS Details

  • SHA256: bddaaec0975f8fa3841f081c6864a05cbc041b0e56aa8c5bcf706257c83c7bb3
  • Pointer size: 130 Bytes
  • Size of remote file: 74.4 kB
scenes/isaac/47331202/20261114T083400Z/input/generation_spec.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "contract": "scenes-isaac",
3
+ "engine": "Isaac Sim",
4
+ "task": "Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.",
5
+ "inputs": [
6
+ "video.mp4",
7
+ "scale_anchor.json",
8
+ "anchor_reference.png",
9
+ "anchor_marked.png",
10
+ "self_check.py"
11
+ ],
12
+ "prompt": "prompt.txt",
13
+ "units": "metres, kilograms, seconds",
14
+ "up_axis": "+Z",
15
+ "entrypoint": "scene.usda",
16
+ "budget": {
17
+ "wall_clock_seconds": 900,
18
+ "scope": "per_scene",
19
+ "starts_at": "agent_input_access",
20
+ "includes_all_tool_and_validation_time": true,
21
+ "deadline_action": "stop and submit latest valid loadable output",
22
+ "enforcement": "Docker generation runner terminates the container at the wall-clock deadline"
23
+ },
24
+ "requirement": "Sim-ready: ready for interactive robot learning, with no manual repair.",
25
+ "sim_ready": {
26
+ "loads_headless_offline": true,
27
+ "resets_to_rest": true,
28
+ "collision_matches_visible_geometry": true,
29
+ "rigid_things": "rigid bodies; free objects movable, articulated parts on limited joints",
30
+ "soft_things": [
31
+ "cloth",
32
+ "bedding",
33
+ "curtains",
34
+ "rugs",
35
+ "throws",
36
+ "cushions",
37
+ "pillows"
38
+ ],
39
+ "soft_mechanism": "native deformable",
40
+ "stable_under_sustained_interaction": true
41
+ },
42
+ "visual_objective": "Maximize visual quality, realism and fidelity to the source video, keeping every sim-ready requirement.",
43
+ "refinement": {
44
+ "minimum_completed_rounds": 3,
45
+ "minimum_reference_views": 6,
46
+ "minimum_inspected_video_frames": 12,
47
+ "final_validation_seconds": 90,
48
+ "outside_data_allowed": false
49
+ },
50
+ "label": "free text saying plainly what the object is; one entry per physical instance",
51
+ "interaction_points": [
52
+ "handle",
53
+ "grasp",
54
+ "spout",
55
+ "material_corners"
56
+ ],
57
+ "construction_rules": {
58
+ "max_initial_penetration_m": 0.005,
59
+ "max_penetration_duration_s": 0.02,
60
+ "max_passive_speed_mps": 5.0,
61
+ "articulated_parts_clear_of_carcass": true,
62
+ "swing_paths_clear": true,
63
+ "timestep": "agent's choice"
64
+ },
65
+ "self_check": "/isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check"
66
+ }
scenes/isaac/47331202/20261114T083400Z/input/prompt.txt ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ About the simulator. NVIDIA Isaac Sim is a robotics simulator built on Omniverse. You describe a scene
2
+ as an OpenUSD stage: prims carrying physics schemas for rigid bodies, colliders, joints and articulations, and
3
+ deformable bodies and surfaces. NVIDIA PhysX simulates it, here with GPU dynamics, and RTX path tracing renders it.
4
+ Isaac Sim is widely used for robot learning and sim-to-real work.
5
+
6
+ Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.
7
+
8
+ THE RESULT MUST BE SIM-READY, READY FOR INTERACTIVE ROBOT LEARNING. It is an environment a robot learns
9
+ in, not a picture of a room. A training pipeline must be able to load it headless, reset it, step it for
10
+ a long time and let robots act on it, with no manual repair. That means:
11
+ - It loads with no network access or external files, and resets to the same state at rest.
12
+ - Everything a robot could touch has collision geometry matching what is visible. An object with no
13
+ collision does not exist for a robot, however good it looks.
14
+ - Rigid things are rigid bodies with plausible mass; free objects can be pushed, and doors, drawers and lids
15
+ move on joints with limits.
16
+ - Soft things are native deformables: cloth, bedding, curtains, rugs, throws, cushions and pillows.
17
+ A rigid body that merely looks soft is wrong.
18
+ - It stays stable under sustained interaction, across many steps without a reset, and steps fast
19
+ enough to train on.
20
+
21
+ Budget: 15 minutes (900 seconds) of wall-clock time per scene, starting when
22
+ you receive access to the input files. All inspection, coding, asset creation,
23
+ tool calls, rendering, debugging, and validation must fit within this budget.
24
+ Stop work at the deadline and submit the latest valid, loadable result.
25
+ Do not stop after producing the first runnable scene. Use the available time
26
+ for repeated self-critique and refinement; reserve the final 90 seconds for
27
+ validation and submission.
28
+
29
+ Inputs (in /work/input):
30
+ - video.mp4: the complete source video.
31
+ - scale_anchor.json: dimensions and identity of one object in the room, the scale anchor.
32
+ - anchor_reference.png and anchor_marked.png: images identifying that object.
33
+ The marked rectangle is a projected box envelope, not a segmentation mask.
34
+ - generation_spec.json: this contract in machine-readable form.
35
+ - self_check.py: the public self-check described below.
36
+
37
+ Maximize visual realism and fidelity to the source video. Reproduce the room's
38
+ geometry, proportions, native objects, materials, textures, colors, lighting,
39
+ and visible details as faithfully as possible within the time budget. Use the
40
+ scale anchor to establish metric scale; infer other dimensions and layout from
41
+ the video. The anchor's known real dimensions let you work out every other
42
+ object's size and the distances between objects, for example by measuring them
43
+ against the anchor in frames where both are visible, or by scaling a
44
+ reconstruction so that the anchor has its true size. Place the anchor object
45
+ where it is in the video; the room is not aligned to it alone.
46
+
47
+ Visual objective: MAXIMIZE THE VISUAL QUALITY of the reconstructed room, without giving up any
48
+ sim-ready requirement: a scene that looks right but cannot be simulated has failed.
49
+ Use source-aligned renders to improve what a viewer actually sees: recognizable
50
+ furniture silhouettes, correct proportions and placement, detailed surfaces,
51
+ plausible texture scale, window appearance, soft lighting and small distinctive
52
+ objects. Avoid a room made only of plain boxes when the video shows richer
53
+ shapes and materials. Create local textures or meshes when they improve fidelity;
54
+ do not use a single photograph or video plane as a substitute for 3D geometry.
55
+ Check the entire camera path, including corners and reverse views, rather than
56
+ optimizing one attractive render.
57
+
58
+ Required self-critique and refinement:
59
+ 1. Inspect at least 12 frames distributed across the video. Choose at least
60
+ six representative reference views covering the room and important objects.
61
+ Record their timestamps. Estimate corresponding cameras using ONLY the
62
+ public video and scale anchor, and reuse those cameras across iterations.
63
+ 2. Build an early loadable baseline and render those views. Do not spend the
64
+ whole budget on a first draft or a single attractive camera angle.
65
+ 3. Complete at least THREE render-critique-refine-verify rounds. In each round,
66
+ open side-by-side source/render contact sheets with the image-capable Read
67
+ tool. Identify concrete visible discrepancies, implement the highest-impact
68
+ fixes, then re-render the SAME views and check for regressions. Writing a
69
+ critique without changing and re-rendering the scene is not a completed round.
70
+ 4. Round 1 prioritizes room shape, scale, furniture placement, object dimensions
71
+ and silhouettes. Round 2 prioritizes materials, textures, lighting, colors,
72
+ windows and salient details. Round 3 revisits visual mismatches and verifies
73
+ collision support, movable parts and soft items, and runs the self-check.
74
+ 5. Continue additional rounds while time permits. Do not wait until the deadline
75
+ to save a valid scene. Report the true number of completed rounds.
76
+ 6. After each round, save a valid checkpoint and record the source timestamps,
77
+ render paths, specific discrepancies, actual edits, before/after observations,
78
+ and remaining issues in refinement_log.json. Treat self-assessments as
79
+ diagnostics.
80
+ 7. Keep the best validated complete scene, and restore it if a later change
81
+ regresses. The final output must describe that selected result.
82
+
83
+ You may use any tools, including web search, and the software installed here,
84
+ including Blender (`blender`). Download nothing from the internet (no files,
85
+ models, assets, datasets or packages), and do not look up this room, its video
86
+ or the dataset it comes from. Use no annotations, scans or other data about this
87
+ room beyond your inputs, no previous generated outputs and no other session's
88
+ reconstruction. Do not read authentication files. Do not add a robot.
89
+
90
+ The result must be a genuinely interactive 3D scene:
91
+ - Preserve native object identities, count, dimensions, and spatial relationships.
92
+ - Provide collision geometry consistent with visible surfaces and free spaces.
93
+ - Separate movable objects and articulated parts, with appropriate joints and limits.
94
+ - Build every visible soft item with the simulator's native deformable mechanism, as one object
95
+ entry each, with a label that says plainly what it is, for example "curtain".
96
+ - Preserve openings, container interiors, support surfaces, and access needed
97
+ for interactions. Do not replace the room's own objects with generic stand-ins.
98
+ - Assign physically plausible properties and document estimates for unknowns;
99
+ the video is not measured mass, friction or stiffness.
100
+
101
+ Construction rules (the self-check reports them):
102
+ - The scene must start at rest without interpenetration: no contact deeper than
103
+ 5 mm between any two bodies or between a body and fixed geometry.
104
+ - Articulated parts (drawers, doors, lids) must not intersect their own carcass in
105
+ any position of their range: hollow the carcass or exclude the pair from collision.
106
+ - Door and drawer swing paths must be clear of other furniture.
107
+ - The timestep is your choice.
108
+
109
+ Object labels: each objects.json entry has a "label" that says plainly what the object is
110
+ (for example "double bed", "floor lamp"). Use one entry per physical instance.
111
+
112
+ Interaction points: for articulated furniture give a "handle" point on each moving
113
+ part; for movable graspable objects a "grasp" point; for pourable containers a
114
+ "spout" point; for a visible bed cover, four "material_corners" in boundary order.
115
+
116
+ Self-check: run it on your output after each round and before submitting. It loads
117
+ only your output, simulates 3 s from the authored state and reports contacts deeper
118
+ than 5 mm (with the two bodies involved), the time above 5 mm, peak body speed,
119
+ non-finite state and entries without a label. The scene must stay within these
120
+ limits: at most 0.02 s above 5 mm, peak speed at most 5 m/s, finite state. Fix what it reports; do not edit it.
121
+ /isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check
122
+
123
+ Output format (Isaac Sim, in /work/output):
124
+ - scene.usda (or scene.usd): self-contained USD stage, Z-up, metres, explicit default
125
+ prim, locally referenced assets and textures, and an authored PhysX scene.
126
+ - build_scene.py: reproducible builder using the installed Isaac Sim / pxr APIs.
127
+ - objects.json: {"objects": [...]}, one entry per physical object instance:
128
+ "id": a persistent unique string (keep the scale anchor ID),
129
+ "label": what the object is, in plain words,
130
+ "prim_path": the object's root prim (no single mesh containing all furniture),
131
+ "joint_paths": articulated joint prims, if present,
132
+ "interaction_points": interaction Xform prim paths, named handle, grasp or spout
133
+ (child Xforms of the moving part or object),
134
+ "material_corners": four mesh vertex indices for a deformable bed cover.
135
+ - render_preview.py, at least six views rendered with RTX PathTracing, preview.mp4
136
+ (camera tour) and interaction.mp4 (native physics interaction with an original object).
137
+ - README.md and refinement_log.json, as above.
138
+ Runtime: /isaac-sim/python.sh YOUR_SCRIPT.py runs the bundled Python with Isaac Sim,
139
+ USD, NumPy and Pillow; /usr/local/bin/ffmpeg and blender are available. Use headless=True,
140
+ multi_gpu=False, active_gpu=0, physics_gpu=0. sdk_smoke.py is an API and rendering
141
+ example. Native RTX capture can need several app updates before RGB has shape
142
+ (height, width, 4). Only input and output are host-mounted. Do not access
143
+ authentication.
scenes/isaac/47331202/20261114T083400Z/input/scale_anchor.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "object_id": "scale_anchor",
3
+ "description": "The table outlined in red in the marked image.",
4
+ "reference_image": "anchor_reference.png",
5
+ "marked_image": "anchor_marked.png",
6
+ "region_xyxy_pixels": [
7
+ 50.18202989863427,
8
+ 9.840496777454728,
9
+ 212.63893423128675,
10
+ 185.5016786254594
11
+ ],
12
+ "region_type": "projected_3d_box_envelope_not_segmentation",
13
+ "size_local_xyz_m": [
14
+ 0.803188210828418,
15
+ 1.314670830811198,
16
+ 0.7819756633824008
17
+ ],
18
+ "axis_description": [
19
+ "horizontal depth",
20
+ "horizontal width",
21
+ "vertical height"
22
+ ],
23
+ "measurement_type": "full extents of an oriented box around the object",
24
+ "uncertainty_m": null,
25
+ "uncertainty_note": "No uncertainty is supplied for these dimensions."
26
+ }
scenes/isaac/47331202/20261114T083400Z/input/sdk_smoke.py ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Verify Isaac Sim headless RTX rendering and native rigid-body stepping."""
2
+ from isaacsim import SimulationApp
3
+ app = SimulationApp({'headless':True,'renderer':'PathTracing','samples_per_pixel_per_frame':32,
4
+ 'width':512,'height':384,'multi_gpu':False,'active_gpu':0,'physics_gpu':0,
5
+ 'enable_crashreporter':False,'limit_cpu_threads':8})
6
+ import json
7
+ from pathlib import Path
8
+ import numpy as np
9
+ from PIL import Image
10
+ import omni.usd
11
+ import omni.replicator.core as rep
12
+ rep.orchestrator.set_capture_on_play(False)
13
+ from pxr import UsdGeom,UsdLux,UsdShade,UsdPhysics,Gf,Sdf
14
+ stage=omni.usd.get_context().get_stage()
15
+ UsdGeom.SetStageUpAxis(stage,UsdGeom.Tokens.z);UsdGeom.SetStageMetersPerUnit(stage,1.)
16
+ UsdGeom.Xform.Define(stage,'/World');stage.SetDefaultPrim(stage.GetPrimAtPath('/World'))
17
+ def box(path,pos,scale,color,dynamic=False):
18
+ prim=UsdGeom.Cube.Define(stage,path);prim.CreateSizeAttr(1.)
19
+ xf=UsdGeom.Xformable(prim);xf.AddTranslateOp().Set(Gf.Vec3d(*pos));xf.AddScaleOp().Set(Gf.Vec3f(*scale))
20
+ UsdPhysics.CollisionAPI.Apply(prim.GetPrim())
21
+ if dynamic:
22
+ UsdPhysics.RigidBodyAPI.Apply(prim.GetPrim());UsdPhysics.MassAPI.Apply(prim.GetPrim()).CreateMassAttr(1.)
23
+ mat=UsdShade.Material.Define(stage,path+'_Material');shader=UsdShade.Shader.Define(stage,path+'_Material/Shader');shader.CreateIdAttr('UsdPreviewSurface')
24
+ shader.CreateInput('diffuseColor',Sdf.ValueTypeNames.Color3f).Set(Gf.Vec3f(*color));shader.CreateInput('roughness',Sdf.ValueTypeNames.Float).Set(.28)
25
+ mat.CreateSurfaceOutput().ConnectToSource(shader.ConnectableAPI(),'surface');UsdShade.MaterialBindingAPI.Apply(prim.GetPrim()).Bind(mat)
26
+ return prim
27
+ box('/World/Floor',(0,0,-.05),(5,5,.1),(.6,.6,.6));box('/World/Cube',(0,0,1.),(.3,.3,.3),(.08,.3,.8),True)
28
+ light=UsdLux.DomeLight.Define(stage,'/World/Dome');light.CreateIntensityAttr(500.)
29
+ light=UsdLux.RectLight.Define(stage,'/World/Key');light.CreateIntensityAttr(2500.);light.CreateWidthAttr(3.);light.CreateHeightAttr(3.);UsdGeom.Xformable(light).AddTranslateOp().Set(Gf.Vec3d(0,0,3))
30
+ scene=UsdPhysics.Scene.Define(stage,'/World/PhysicsScene');scene.CreateGravityDirectionAttr(Gf.Vec3f(0,0,-1));scene.CreateGravityMagnitudeAttr(9.81)
31
+ camera=rep.create.camera(position=(2.4,-3,2),look_at=(0,0,.3));product=rep.create.render_product(camera,(512,384));rgb=rep.AnnotatorRegistry.get_annotator('rgb');rgb.attach([product])
32
+ from isaacsim.core.api import World
33
+ world=World(physics_dt=1/60,rendering_dt=1/60,stage_units_in_meters=1.)
34
+ world.reset()
35
+ for _ in range(30):world.step(render=True)
36
+ for attempt in range(30):
37
+ rep.orchestrator.step(rt_subframes=8)
38
+ app.update()
39
+ image=np.asarray(rgb.get_data())
40
+ print('CAPTURE_ATTEMPT',attempt,'SHAPE',image.shape,flush=True)
41
+ if image.ndim==3 and image.shape[0]>0:break
42
+ print('RGB_SHAPE',image.shape,flush=True)
43
+ assert image.ndim==3,'RTX annotator did not return a frame'
44
+ position=UsdGeom.Xformable(stage.GetPrimAtPath('/World/Cube')).ComputeLocalToWorldTransform(0).ExtractTranslation()
45
+ print('CUBE_POSITION',list(position),flush=True)
46
+ out=Path('/work/output');out.mkdir(exist_ok=True)
47
+ Image.fromarray(image[:,:,:3]).save(out/'smoke.png');stage.GetRootLayer().Export(str(out/'sdk_example.usda'))
48
+ (out/'smoke.json').write_text(json.dumps({'renderer':'Isaac Sim RTX PathTracing','shape':list(image.shape),'cube_final_position':list(position),'initial_height_m':1.,'success':True}))
49
+ app.close()
scenes/isaac/47331202/20261114T083400Z/input/self_check.py ADDED
@@ -0,0 +1,228 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Passive physics of a USD scene in Isaac Sim, against the scene requirements (no penetration beyond 5 mm for more than
2
+ 0.02 s, no body faster than 5 m/s, finite state).
3
+
4
+ Runs inside the Isaac Sim container. Reads --scene (default /work/scene, read-only) and the optional
5
+ --out/request.json (object ids to drop); writes --out/physics.json (default /work/physics). Without a request it is
6
+ the self-check: no drops, plus a check that every object has a non-empty label.
7
+
8
+ - Speed: a 3 s passive rollout with no rendering and no state reads, timed around world.step only
9
+ (simulated seconds / wall seconds).
10
+ - Passive rollout: 3 s from the authored state. Penetration comes from the PhysX tensor contact view
11
+ (GPU contact reports return nothing under GPU dynamics); speed and poses from the RigidPrim view.
12
+ - Drops: the requested objects with free (unjointed) dynamic bodies, raised 0.05 m
13
+ and released for 2 s, each from a fresh reset.
14
+ The limits are the same as for MuJoCo scenes.
15
+ """
16
+ from isaacsim import SimulationApp
17
+ app=SimulationApp({'headless':True,'multi_gpu':False,'active_gpu':0,'physics_gpu':0,
18
+ 'enable_crashreporter':False,'limit_cpu_threads':8})
19
+ import argparse
20
+ import json
21
+ from pathlib import Path
22
+ import time
23
+ import carb
24
+ import numpy as np
25
+ import omni.usd
26
+ from pxr import UsdPhysics,PhysxSchema
27
+
28
+ LIMIT_M,LIMIT_S,SPEED=.005,.02,5.
29
+ parser=argparse.ArgumentParser();parser.add_argument('--scene',default='/work/scene');parser.add_argument('--out',default='/work/physics')
30
+ args,_=parser.parse_known_args()
31
+ out=Path(args.out);out.mkdir(parents=True,exist_ok=True)
32
+ request=json.loads((out/'request.json').read_text()) if (out/'request.json').is_file() else {'object_ids':[]}
33
+ result={'status':'failed','engine':'Isaac Sim PhysX (GPU dynamics)','no_vlm':True,'provisional':True,
34
+ 'thresholds':{'penetration_m':LIMIT_M,'penetration_duration_s':LIMIT_S,'linear_speed_mps':SPEED},
35
+ 'thresholds_calibrated_across_engines':False}
36
+
37
+
38
+ def contact_filters(stage,paths):
39
+ """GPU contact views filter on rigid bodies or standalone colliders only, so a collider inside a
40
+ rigid body is replaced by that body (unsupported filters are dropped with only a warning)."""
41
+ out=[]
42
+ for path in paths:
43
+ if any('deformable' in str(a).lower() for a in stage.GetPrimAtPath(path).GetAppliedSchemas()):
44
+ continue # deformable meshes cannot be contact filters; their contacts are not measured
45
+ prim=stage.GetPrimAtPath(path);owner=None
46
+ while prim and not prim.IsPseudoRoot():
47
+ if prim.HasAPI(UsdPhysics.RigidBodyAPI):owner=str(prim.GetPath())
48
+ prim=prim.GetParent()
49
+ target=owner or path
50
+ if target not in out:out.append(target)
51
+ return out
52
+
53
+
54
+ def array(value):
55
+ return np.asarray(value.cpu() if hasattr(value,'cpu') else value,dtype=float)
56
+
57
+
58
+ try:
59
+ root=Path(args.scene)
60
+ scene=next(root/n for n in ['scene.usda','scene.usd'] if (root/n).is_file())
61
+ omni.usd.get_context().open_stage(str(scene));stage=omni.usd.get_context().get_stage()
62
+ for _ in range(8):app.update()
63
+
64
+ def dynamic(prim):
65
+ api=UsdPhysics.RigidBodyAPI(prim)
66
+ if api.GetRigidBodyEnabledAttr().Get() not in (None,True) or api.GetKinematicEnabledAttr().Get():return False
67
+ parent=prim.GetParent()
68
+ while parent and not parent.IsPseudoRoot():
69
+ if parent.HasAPI(UsdPhysics.RigidBodyAPI):return False
70
+ parent=parent.GetParent()
71
+ return True
72
+ rigid=[p for p in stage.Traverse() if p.HasAPI(UsdPhysics.RigidBodyAPI)]
73
+ bodies=[p for p in rigid if dynamic(p)]
74
+ paths=[str(p.GetPath()) for p in bodies]
75
+ colliders=[str(p.GetPath()) for p in stage.Traverse() if p.HasAPI(UsdPhysics.CollisionAPI)]
76
+ jointed=set()
77
+ for prim in stage.Traverse():
78
+ if prim.IsA(UsdPhysics.Joint):
79
+ joint=UsdPhysics.Joint(prim)
80
+ for rel in (joint.GetBody0Rel(),joint.GetBody1Rel()):
81
+ jointed.update(str(t) for t in rel.GetTargets())
82
+ for prim in bodies:PhysxSchema.PhysxContactReportAPI.Apply(prim).CreateThresholdAttr().Set(0.)
83
+ scenes=[p for p in stage.Traverse() if p.IsA(UsdPhysics.Scene)]
84
+ authored=PhysxSchema.PhysxSceneAPI(scenes[0]).GetTimeStepsPerSecondAttr().Get() if scenes else None
85
+ rate=float(authored or 60)
86
+ result.update(dynamic_bodies=len(paths),colliders=len(colliders),
87
+ excluded_bodies=[str(p.GetPath()) for p in rigid if not dynamic(p)],
88
+ timestep_s=1/rate,timestep_source='authored PhysxSceneAPI timeStepsPerSecond' if authored else 'default 60 Hz')
89
+ if not paths:raise ValueError('No dynamic rigid bodies to simulate')
90
+ filters=contact_filters(stage,colliders)
91
+ result['contact_filters']=len(filters)
92
+
93
+ carb.settings.get_settings().set_bool('/physics/disableContactProcessing',False)
94
+ from isaacsim.core.api import World
95
+ from isaacsim.core.prims import RigidPrim
96
+ from isaacsim.core.simulation_manager import SimulationManager
97
+ world=World(physics_dt=1/rate,rendering_dt=1/rate,stage_units_in_meters=1.,
98
+ physics_prim_path=str(scenes[0].GetPath()) if scenes else '/World/AddedPhysicsScene',
99
+ set_defaults=False,backend='torch',device='cuda:0')
100
+ context=world.get_physics_context()
101
+ context.enable_gpu_dynamics(True);context.set_broadphase_type('GPU')
102
+ view=RigidPrim(paths,name='scene_bodies');world.scene.add(view)
103
+ steps=lambda seconds:int(np.ceil(seconds*rate))
104
+
105
+ def measure_rtf(dt,repeats=3):
106
+ """Simulated seconds / wall seconds of world.step(render=False) for an unread passive rollout from the authored
107
+ state, median over `repeats` after 10 warm-up steps."""
108
+ world.set_simulation_dt(physics_dt=dt,rendering_dt=dt)
109
+ world.reset()
110
+ for _ in range(10):world.step(render=False)
111
+ count=int(np.ceil(3/dt));samples=[];finite=True
112
+ for _ in range(repeats):
113
+ world.reset();start=time.perf_counter()
114
+ for _ in range(count):world.step(render=False)
115
+ samples.append(count*dt/(time.perf_counter()-start))
116
+ positions,_=view.get_world_poses()
117
+ finite&=bool(np.isfinite(array(positions)).all() and np.isfinite(array(view.get_velocities())).all())
118
+ return {'rtf':float(np.median(samples)),'rtf_samples':samples,'timestep_s':dt,'simulated_seconds':count*dt,
119
+ 'steps':count,'repeats':repeats,'finite':finite,
120
+ 'definition':'simulated seconds / wall seconds of world.step(render=False) only, median of repeats'}
121
+
122
+ result['realtime']={**measure_rtf(1/rate),'scope':'3 s passive rollout from the authored state, authored timestep'}
123
+ print('PHYSICS_STAGE realtime',result['realtime']['rtf'],flush=True)
124
+
125
+ def reset():
126
+ world.reset()
127
+ contacts=SimulationManager.get_physics_sim_view().create_rigid_contact_view(
128
+ paths,filter_patterns=[filters]*len(paths),max_contact_data_count=200000)
129
+ if contacts is None:raise RuntimeError('PhysX rigid contact view could not be created')
130
+ return contacts
131
+
132
+ def rollout(seconds,contacts,motion=None):
133
+ """One rollout; `motion` names the file the logged poses are written to."""
134
+ report={'finite':True,'peak_penetration_m':0.,'penetration_time_s':0.,'peak_linear_speed_mps':0.}
135
+ every=max(1,round(rate/10));frames=[]
136
+ for step in range(steps(seconds)):
137
+ world.step(render=False)
138
+ positions,orientations=view.get_world_poses();velocities=array(view.get_velocities())
139
+ if step%every==0:frames.append(((step+1)/rate,array(positions),array(orientations)))
140
+ report['finite']&=bool(np.isfinite(array(positions)).all() and np.isfinite(velocities).all())
141
+ _,_,_,separation,counts,starts=contacts.get_contact_data(dt=1/rate)
142
+ total=int(array(counts).sum())
143
+ if 'initial_pairs' not in report:
144
+ # Pairs deeper than the limit after the first step, named for the self-check.
145
+ pairs=[];sep=array(separation);cnt=array(counts).astype(int);beg=array(starts).astype(int)
146
+ for i,j in zip(*np.nonzero(cnt)):
147
+ depth=-float(sep[beg[i,j]:beg[i,j]+cnt[i,j]].min())
148
+ if depth>LIMIT_M:pairs.append({'depth_m':depth,'bodies':[paths[i],filters[j]]})
149
+ report['initial_pairs']=sorted(pairs,key=lambda p:-p['depth_m'])[:20]
150
+ deepest=max(0.,-float(array(separation)[:total].min())) if total else 0.
151
+ report['peak_penetration_m']=max(report['peak_penetration_m'],deepest)
152
+ if deepest>LIMIT_M:report['penetration_time_s']+=1/rate
153
+ report['peak_linear_speed_mps']=max(report['peak_linear_speed_mps'],float(np.linalg.norm(velocities[:,:3],axis=1).max()))
154
+ if not report['finite']:break
155
+ report['valid']=bool(report['finite'] and report['penetration_time_s']<=LIMIT_S and report['peak_linear_speed_mps']<=SPEED)
156
+ report['thresholds']=result['thresholds']
157
+ if motion and request.get('record'):
158
+ (out/'motion').mkdir(exist_ok=True)
159
+ np.savez_compressed(out/'motion'/motion,paths=np.array(paths),t=np.array([f[0] for f in frames]),
160
+ positions=np.stack([f[1] for f in frames]),orientations=np.stack([f[2] for f in frames]))
161
+ report['motion']=f'motion/{motion}.npz'
162
+ return report
163
+
164
+ contacts=reset()
165
+ start_positions=array(view.get_world_poses()[0])
166
+ masses=array(view.get_masses())
167
+ direction,magnitude=context.get_gravity()
168
+ gravity=np.asarray(direction,dtype=float)*float(magnitude)
169
+ passive=rollout(3,contacts,'passive')
170
+ print('PHYSICS_STAGE passive',passive,flush=True)
171
+ end_positions=array(view.get_world_poses()[0])
172
+ moved=np.linalg.norm(end_positions-start_positions,axis=1)
173
+
174
+ manifest=json.loads((root/'objects.json').read_text())
175
+ manifest=manifest['objects'] if isinstance(manifest,dict) else manifest
176
+ result['manifest_errors']=[f"{o.get('id')}: label is missing or empty" for o in manifest if not str(o.get('label') or '').strip()]
177
+ result['manifest_errors']+=[f"{o.get('id')}: prim_path does not resolve" for o in manifest if not stage.GetPrimAtPath(o.get('prim_path','') or '/')]
178
+ def owned(obj):
179
+ prefix=str(obj.get('prim_path','')).rstrip('/')
180
+ return [i for i,p in enumerate(paths) if prefix and (p==prefix or p.startswith(prefix+'/'))]
181
+ drift=[{'id':o['id'],'dynamic_bodies':len(owned(o)),'center_drift_m':float(moved[owned(o)].max()),'mass_kg':float(masses[owned(o)].sum())}
182
+ for o in manifest if owned(o)]
183
+
184
+ requested=set(request['object_ids'])
185
+ eligible=[o for o in manifest if o['id'] in requested]
186
+ drops=[]
187
+ for obj in eligible:
188
+ free=[i for i in owned(obj) if paths[i] not in jointed]
189
+ if not free:continue
190
+ contacts=reset()
191
+ positions,orientations=view.get_world_poses()
192
+ raised=positions.clone() if hasattr(positions,'clone') else positions.copy()
193
+ raised[free,2]+=.05
194
+ view.set_world_poses(positions=raised,orientations=orientations)
195
+ applied=float((array(view.get_world_poses()[0])[free,2]-array(positions)[free,2]).min())
196
+ drops.append({'id':obj['id'],'input':'raise free bodies 0.05 m and release','bodies':[paths[i] for i in free],
197
+ 'lift_applied_m':applied,**rollout(2,contacts,f'drop_{len(drops)}')})
198
+ print('PHYSICS_STAGE drop',obj['id'],applied,drops[-1]['peak_linear_speed_mps'],flush=True)
199
+
200
+ # The common 2 ms step last, so the rollouts above run at the authored step.
201
+ result['realtime_common_step']={**measure_rtf(.002),'scope':'3 s passive rollout from the authored state at the common 2 ms step'}
202
+ print('PHYSICS_STAGE realtime_common_step',result['realtime_common_step']['rtf'],flush=True)
203
+ gravity_ok=bool(np.allclose(gravity,[0,0,-9.81],atol=.1))
204
+ mass_ok=bool(np.isfinite(masses).all() and (masses>=0).all())
205
+ result.update(status='completed',passive=passive,gravity_m_s2=gravity.tolist(),gravity_valid=gravity_ok,
206
+ mass_valid=mass_ok,center_drift=drift,drop_tests=drops,eligible_native_objects=len(eligible),
207
+ largest_body_displacements=[{'prim':paths[i],'displacement_m':float(moved[i])} for i in np.argsort(-moved)[:5]],
208
+ valid=bool(passive['valid'] and gravity_ok and mass_ok),
209
+ drop_scope='the requested objects with free (unjointed) dynamic bodies, in manifest order',
210
+ contact_scope=('contacts between each dynamic body and every collider (static, kinematic and dynamic; colliders '
211
+ 'inside a rigid body are filtered by that body, as GPU contact views require); '
212
+ 'static-static pairs are not simulated, as in MuJoCo world-body geoms'),
213
+ limitations=['the limits are the MuJoCo ones, not calibrated for PhysX',
214
+ 'kinematic, disabled and nested rigid bodies are not traced (see excluded_bodies)',
215
+ 'drift is per dynamic body under the object prim, not the object bounding box',
216
+ 'contacts with deformable bodies are not measured (GPU contact views cannot filter them)'])
217
+ except Exception as error:
218
+ result['error']=f'{type(error).__name__}: {error}'
219
+ finally:
220
+ (out/'physics.json').write_text(json.dumps(result,indent=2)+'\n')
221
+ if 'passive' in result:
222
+ passive=result['passive']
223
+ print('PASSIVE CHECK','PASSED' if passive['valid'] and not result.get('manifest_errors') else 'FAILED',
224
+ f"- {len(passive.get('initial_pairs',[]))} contact pairs > 5 mm, {passive['penetration_time_s']:.3f} s above 5 mm,"
225
+ f" peak speed {passive['peak_linear_speed_mps']:.2f} m/s, {len(result.get('manifest_errors',[]))} manifest errors;"
226
+ f" details in {out/'physics.json'}",flush=True)
227
+ else:print('CHECK FAILED:',result.get('error'),flush=True)
228
+ app.close()
scenes/isaac/47331202/20261114T083400Z/input/video.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:066fcbfb5a6f9ffecd6af6845e654de29195adc2c3d233a9092a11c8b81c0e2a
3
+ size 11908131
scenes/isaac/47331202/20261114T083400Z/run.json ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "track": "scenes",
3
+ "engine": "isaac",
4
+ "item": "47331202",
5
+ "run_id": "20261114T083400Z",
6
+ "campaign": "jiarui-2-codebuddy",
7
+ "cli": "codebuddy",
8
+ "model": "Hunyuan 4",
9
+ "model_id": "hy4-preview",
10
+ "model_family": "hy4-preview",
11
+ "model_reasoning_effort": "default",
12
+ "budget_seconds": 900,
13
+ "told_budget_seconds": 900,
14
+ "stage": 1,
15
+ "continued_from": null,
16
+ "contract": {
17
+ "name": "scenes-isaac",
18
+ "sha256": "f1ab7ea88842502d37176b41d9f3bec64a5d9fac292ca3f16caf08da3b5db96d",
19
+ "track": "scenes",
20
+ "engine": "isaac",
21
+ "engine_version": "6.2.0",
22
+ "image": "simready-isaac-generation:latest",
23
+ "blender_version": "5.2.3",
24
+ "inputs": [
25
+ "video.mp4",
26
+ "scale_anchor.json",
27
+ "anchor_reference.png",
28
+ "anchor_marked.png",
29
+ "prompt.txt",
30
+ "generation_spec.json",
31
+ "self_check.py",
32
+ "sdk_smoke.py"
33
+ ],
34
+ "withheld": [
35
+ "groundtruth/ (objects, cameras, task catalogs, object dynamics, interaction test) and its source tag",
36
+ "provenance/ (source annotation, anchor selection)",
37
+ "the frozen protocol of the track and engine (data/benchmark/frozen/<track>/<engine>.json)",
38
+ "other runs"
39
+ ],
40
+ "outputs": [
41
+ "scene.usda",
42
+ "build_scene.py",
43
+ "objects.json",
44
+ "render_preview.py",
45
+ "preview.mp4",
46
+ "interaction.mp4",
47
+ "README.md",
48
+ "refinement_log.json"
49
+ ],
50
+ "self_check": "/isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check",
51
+ "budget": {
52
+ "told": "the first of the session's cumulative budgets, in wall-clock seconds from container start",
53
+ "resume": "each later budget resumes the same conversation for the rest of its time; each budget is its own run",
54
+ "enforcement": "the runner kills the container at the deadline; the output as it stands is the result"
55
+ },
56
+ "refinement": {
57
+ "minimum_completed_rounds": 3,
58
+ "minimum_reference_views": 6,
59
+ "minimum_inspected_video_frames": 12,
60
+ "final_validation_seconds": 90
61
+ },
62
+ "isolation": {
63
+ "container": "Docker: read-only input (and tools), writable output and scratch, temporary credentials; no repository, benchmark ground truth or evaluator files",
64
+ "network": "an internal network whose only exit is an HTTPS CONNECT proxy (runtime/egress_proxy.py)",
65
+ "egress_hosts": [
66
+ "api.anthropic.com",
67
+ "platform.claude.com",
68
+ "chatgpt.com",
69
+ "auth.openai.com",
70
+ "opencode.ai",
71
+ "mcp.exa.ai",
72
+ "search.parallel.ai"
73
+ ]
74
+ },
75
+ "protocol": {
76
+ "file": "data/benchmark/frozen/scenes/isaac.json",
77
+ "sha256": "df0fe5bb1c796474ad13cdba81b660a2946aa24297cb289f9581cfde7c035bb7"
78
+ }
79
+ },
80
+ "engine_version": "6.2.0",
81
+ "image": "simready-isaac-generation:latest",
82
+ "image_id": "sha256:8d9f9c934457aa3c8e91e8099e35db27b50db287d0eb76a647052c310e754d87",
83
+ "required_refinement_rounds": 3,
84
+ "gpu_uuid": "GPU-ab037b41-847d-64be-b2b8-ba682a201943",
85
+ "egress_mode": "internal_network",
86
+ "input_snapshot": "input",
87
+ "tools": null,
88
+ "evaluator_protocol": "evaluator_protocol.json",
89
+ "output": "output",
90
+ "log": "generation.jsonl",
91
+ "status": "budget_expired",
92
+ "agent_cli": {
93
+ "cli": "codebuddy",
94
+ "name": "CodeBuddy Code",
95
+ "binary": "a pinned codebuddy-code package and its node, mounted read-only (SIMREADY_CODEBUDDY_PREFIX, default the host's codebuddy install)",
96
+ "login": "the CodeBuddy login file, given to the CLI as a snapshot without its refresh token (the session cannot rotate it)",
97
+ "billing": "CodeBuddy account (the CLI reports no price)",
98
+ "tools": "every tool but WebFetch and the server-side media generators ImageGen, ImageEdit, VideoGen (--disallowedTools); WebSearch runs on CodeBuddy's side",
99
+ "hosts": [
100
+ "[redacted-host]",
101
+ "wb.[redacted]buddy.com"
102
+ ],
103
+ "version": "2.182.0",
104
+ "model": "hy4-preview",
105
+ "model_family": "hy4-preview"
106
+ },
107
+ "started_at": "2026-11-14T08:34:02.133047+00:00",
108
+ "model_answered": true,
109
+ "exit_code": 137,
110
+ "credential_refresh_persistence": "snapshot_without_refresh_token",
111
+ "finished_at": "2026-11-14T08:49:03.559011+00:00",
112
+ "cli_session_id": "01a1d931-3a32-7b0d-acce-07e307a1a176",
113
+ "api_retries": 0,
114
+ "rate_limit_events": 0,
115
+ "stage_seconds": 901.456,
116
+ "elapsed_seconds": 901.456,
117
+ "cost": {
118
+ "status": "partial",
119
+ "source": "the CLI's running usage totals (usage_update events)",
120
+ "usd": null,
121
+ "usd_basis": "not reported (CodeBuddy account billing)",
122
+ "credits": 0,
123
+ "credit_status": "known",
124
+ "turns": null,
125
+ "api_seconds": null,
126
+ "tokens": {
127
+ "input": 46524,
128
+ "cache_read": 358976,
129
+ "cache_write": 0,
130
+ "output": 6186,
131
+ "reasoning": null,
132
+ "total": 411686
133
+ },
134
+ "wall_seconds": 901.456,
135
+ "budget_seconds": 900,
136
+ "stopped_at_budget": true
137
+ },
138
+ "scene_present": false
139
+ }
scenes/isaac/47331202/20261114T083400Z/submitted_prompt.txt ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ About the simulator. NVIDIA Isaac Sim is a robotics simulator built on Omniverse. You describe a scene
2
+ as an OpenUSD stage: prims carrying physics schemas for rigid bodies, colliders, joints and articulations, and
3
+ deformable bodies and surfaces. NVIDIA PhysX simulates it, here with GPU dynamics, and RTX path tracing renders it.
4
+ Isaac Sim is widely used for robot learning and sim-to-real work.
5
+
6
+ Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.
7
+
8
+ THE RESULT MUST BE SIM-READY, READY FOR INTERACTIVE ROBOT LEARNING. It is an environment a robot learns
9
+ in, not a picture of a room. A training pipeline must be able to load it headless, reset it, step it for
10
+ a long time and let robots act on it, with no manual repair. That means:
11
+ - It loads with no network access or external files, and resets to the same state at rest.
12
+ - Everything a robot could touch has collision geometry matching what is visible. An object with no
13
+ collision does not exist for a robot, however good it looks.
14
+ - Rigid things are rigid bodies with plausible mass; free objects can be pushed, and doors, drawers and lids
15
+ move on joints with limits.
16
+ - Soft things are native deformables: cloth, bedding, curtains, rugs, throws, cushions and pillows.
17
+ A rigid body that merely looks soft is wrong.
18
+ - It stays stable under sustained interaction, across many steps without a reset, and steps fast
19
+ enough to train on.
20
+
21
+ Budget: 15 minutes (900 seconds) of wall-clock time per scene, starting when
22
+ you receive access to the input files. All inspection, coding, asset creation,
23
+ tool calls, rendering, debugging, and validation must fit within this budget.
24
+ Stop work at the deadline and submit the latest valid, loadable result.
25
+ Do not stop after producing the first runnable scene. Use the available time
26
+ for repeated self-critique and refinement; reserve the final 90 seconds for
27
+ validation and submission.
28
+
29
+ Inputs (in /work/input):
30
+ - video.mp4: the complete source video.
31
+ - scale_anchor.json: dimensions and identity of one object in the room, the scale anchor.
32
+ - anchor_reference.png and anchor_marked.png: images identifying that object.
33
+ The marked rectangle is a projected box envelope, not a segmentation mask.
34
+ - generation_spec.json: this contract in machine-readable form.
35
+ - self_check.py: the public self-check described below.
36
+
37
+ Maximize visual realism and fidelity to the source video. Reproduce the room's
38
+ geometry, proportions, native objects, materials, textures, colors, lighting,
39
+ and visible details as faithfully as possible within the time budget. Use the
40
+ scale anchor to establish metric scale; infer other dimensions and layout from
41
+ the video. The anchor's known real dimensions let you work out every other
42
+ object's size and the distances between objects, for example by measuring them
43
+ against the anchor in frames where both are visible, or by scaling a
44
+ reconstruction so that the anchor has its true size. Place the anchor object
45
+ where it is in the video; the room is not aligned to it alone.
46
+
47
+ Visual objective: MAXIMIZE THE VISUAL QUALITY of the reconstructed room, without giving up any
48
+ sim-ready requirement: a scene that looks right but cannot be simulated has failed.
49
+ Use source-aligned renders to improve what a viewer actually sees: recognizable
50
+ furniture silhouettes, correct proportions and placement, detailed surfaces,
51
+ plausible texture scale, window appearance, soft lighting and small distinctive
52
+ objects. Avoid a room made only of plain boxes when the video shows richer
53
+ shapes and materials. Create local textures or meshes when they improve fidelity;
54
+ do not use a single photograph or video plane as a substitute for 3D geometry.
55
+ Check the entire camera path, including corners and reverse views, rather than
56
+ optimizing one attractive render.
57
+
58
+ Required self-critique and refinement:
59
+ 1. Inspect at least 12 frames distributed across the video. Choose at least
60
+ six representative reference views covering the room and important objects.
61
+ Record their timestamps. Estimate corresponding cameras using ONLY the
62
+ public video and scale anchor, and reuse those cameras across iterations.
63
+ 2. Build an early loadable baseline and render those views. Do not spend the
64
+ whole budget on a first draft or a single attractive camera angle.
65
+ 3. Complete at least THREE render-critique-refine-verify rounds. In each round,
66
+ open side-by-side source/render contact sheets with the image-capable Read
67
+ tool. Identify concrete visible discrepancies, implement the highest-impact
68
+ fixes, then re-render the SAME views and check for regressions. Writing a
69
+ critique without changing and re-rendering the scene is not a completed round.
70
+ 4. Round 1 prioritizes room shape, scale, furniture placement, object dimensions
71
+ and silhouettes. Round 2 prioritizes materials, textures, lighting, colors,
72
+ windows and salient details. Round 3 revisits visual mismatches and verifies
73
+ collision support, movable parts and soft items, and runs the self-check.
74
+ 5. Continue additional rounds while time permits. Do not wait until the deadline
75
+ to save a valid scene. Report the true number of completed rounds.
76
+ 6. After each round, save a valid checkpoint and record the source timestamps,
77
+ render paths, specific discrepancies, actual edits, before/after observations,
78
+ and remaining issues in refinement_log.json. Treat self-assessments as
79
+ diagnostics.
80
+ 7. Keep the best validated complete scene, and restore it if a later change
81
+ regresses. The final output must describe that selected result.
82
+
83
+ You may use any tools, including web search, and the software installed here,
84
+ including Blender (`blender`). Download nothing from the internet (no files,
85
+ models, assets, datasets or packages), and do not look up this room, its video
86
+ or the dataset it comes from. Use no annotations, scans or other data about this
87
+ room beyond your inputs, no previous generated outputs and no other session's
88
+ reconstruction. Do not read authentication files. Do not add a robot.
89
+
90
+ The result must be a genuinely interactive 3D scene:
91
+ - Preserve native object identities, count, dimensions, and spatial relationships.
92
+ - Provide collision geometry consistent with visible surfaces and free spaces.
93
+ - Separate movable objects and articulated parts, with appropriate joints and limits.
94
+ - Build every visible soft item with the simulator's native deformable mechanism, as one object
95
+ entry each, with a label that says plainly what it is, for example "curtain".
96
+ - Preserve openings, container interiors, support surfaces, and access needed
97
+ for interactions. Do not replace the room's own objects with generic stand-ins.
98
+ - Assign physically plausible properties and document estimates for unknowns;
99
+ the video is not measured mass, friction or stiffness.
100
+
101
+ Construction rules (the self-check reports them):
102
+ - The scene must start at rest without interpenetration: no contact deeper than
103
+ 5 mm between any two bodies or between a body and fixed geometry.
104
+ - Articulated parts (drawers, doors, lids) must not intersect their own carcass in
105
+ any position of their range: hollow the carcass or exclude the pair from collision.
106
+ - Door and drawer swing paths must be clear of other furniture.
107
+ - The timestep is your choice.
108
+
109
+ Object labels: each objects.json entry has a "label" that says plainly what the object is
110
+ (for example "double bed", "floor lamp"). Use one entry per physical instance.
111
+
112
+ Interaction points: for articulated furniture give a "handle" point on each moving
113
+ part; for movable graspable objects a "grasp" point; for pourable containers a
114
+ "spout" point; for a visible bed cover, four "material_corners" in boundary order.
115
+
116
+ Self-check: run it on your output after each round and before submitting. It loads
117
+ only your output, simulates 3 s from the authored state and reports contacts deeper
118
+ than 5 mm (with the two bodies involved), the time above 5 mm, peak body speed,
119
+ non-finite state and entries without a label. The scene must stay within these
120
+ limits: at most 0.02 s above 5 mm, peak speed at most 5 m/s, finite state. Fix what it reports; do not edit it.
121
+ /isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check
122
+
123
+ Output format (Isaac Sim, in /work/output):
124
+ - scene.usda (or scene.usd): self-contained USD stage, Z-up, metres, explicit default
125
+ prim, locally referenced assets and textures, and an authored PhysX scene.
126
+ - build_scene.py: reproducible builder using the installed Isaac Sim / pxr APIs.
127
+ - objects.json: {"objects": [...]}, one entry per physical object instance:
128
+ "id": a persistent unique string (keep the scale anchor ID),
129
+ "label": what the object is, in plain words,
130
+ "prim_path": the object's root prim (no single mesh containing all furniture),
131
+ "joint_paths": articulated joint prims, if present,
132
+ "interaction_points": interaction Xform prim paths, named handle, grasp or spout
133
+ (child Xforms of the moving part or object),
134
+ "material_corners": four mesh vertex indices for a deformable bed cover.
135
+ - render_preview.py, at least six views rendered with RTX PathTracing, preview.mp4
136
+ (camera tour) and interaction.mp4 (native physics interaction with an original object).
137
+ - README.md and refinement_log.json, as above.
138
+ Runtime: /isaac-sim/python.sh YOUR_SCRIPT.py runs the bundled Python with Isaac Sim,
139
+ USD, NumPy and Pillow; /usr/local/bin/ffmpeg and blender are available. Use headless=True,
140
+ multi_gpu=False, active_gpu=0, physics_gpu=0. sdk_smoke.py is an API and rendering
141
+ example. Native RTX capture can need several app updates before RGB has shape
142
+ (height, width, 4). Only input and output are host-mounted. Do not access
143
+ authentication.
scenes/isaac/47331524/20261114T061812Z/evaluator_protocol.json ADDED
The diff for this file is too large to render. See raw diff
 
scenes/isaac/47331524/20261114T061812Z/generation.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
scenes/isaac/47331524/20261114T061812Z/input/anchor_marked.png ADDED

Git LFS Details

  • SHA256: 2ea99bbde1d2d77e45fc55e41276f0edb0d74cebe5e7a0dd8d09c83be372847a
  • Pointer size: 130 Bytes
  • Size of remote file: 74 kB
scenes/isaac/47331524/20261114T061812Z/input/anchor_reference.png ADDED

Git LFS Details

  • SHA256: fed846168acd61f0d575c69a7d522468bbb28d5e57a326f920055dcfd6224f05
  • Pointer size: 130 Bytes
  • Size of remote file: 74.6 kB
scenes/isaac/47331524/20261114T061812Z/input/generation_spec.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "contract": "scenes-isaac",
3
+ "engine": "Isaac Sim",
4
+ "task": "Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.",
5
+ "inputs": [
6
+ "video.mp4",
7
+ "scale_anchor.json",
8
+ "anchor_reference.png",
9
+ "anchor_marked.png",
10
+ "self_check.py"
11
+ ],
12
+ "prompt": "prompt.txt",
13
+ "units": "metres, kilograms, seconds",
14
+ "up_axis": "+Z",
15
+ "entrypoint": "scene.usda",
16
+ "budget": {
17
+ "wall_clock_seconds": 1800,
18
+ "scope": "per_scene",
19
+ "starts_at": "agent_input_access",
20
+ "includes_all_tool_and_validation_time": true,
21
+ "deadline_action": "stop and submit latest valid loadable output",
22
+ "enforcement": "Docker generation runner terminates the container at the wall-clock deadline"
23
+ },
24
+ "requirement": "Sim-ready: ready for interactive robot learning, with no manual repair.",
25
+ "sim_ready": {
26
+ "loads_headless_offline": true,
27
+ "resets_to_rest": true,
28
+ "collision_matches_visible_geometry": true,
29
+ "rigid_things": "rigid bodies; free objects movable, articulated parts on limited joints",
30
+ "soft_things": [
31
+ "cloth",
32
+ "bedding",
33
+ "curtains",
34
+ "rugs",
35
+ "throws",
36
+ "cushions",
37
+ "pillows"
38
+ ],
39
+ "soft_mechanism": "native deformable",
40
+ "stable_under_sustained_interaction": true
41
+ },
42
+ "visual_objective": "Maximize visual quality, realism and fidelity to the source video, keeping every sim-ready requirement.",
43
+ "refinement": {
44
+ "minimum_completed_rounds": 3,
45
+ "minimum_reference_views": 6,
46
+ "minimum_inspected_video_frames": 12,
47
+ "final_validation_seconds": 90,
48
+ "outside_data_allowed": false
49
+ },
50
+ "label": "free text saying plainly what the object is; one entry per physical instance",
51
+ "interaction_points": [
52
+ "handle",
53
+ "grasp",
54
+ "spout",
55
+ "material_corners"
56
+ ],
57
+ "construction_rules": {
58
+ "max_initial_penetration_m": 0.005,
59
+ "max_penetration_duration_s": 0.02,
60
+ "max_passive_speed_mps": 5.0,
61
+ "articulated_parts_clear_of_carcass": true,
62
+ "swing_paths_clear": true,
63
+ "timestep": "agent's choice"
64
+ },
65
+ "self_check": "/isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check"
66
+ }
scenes/isaac/47331524/20261114T061812Z/input/prompt.txt ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ About the simulator. NVIDIA Isaac Sim is a robotics simulator built on Omniverse. You describe a scene
2
+ as an OpenUSD stage: prims carrying physics schemas for rigid bodies, colliders, joints and articulations, and
3
+ deformable bodies and surfaces. NVIDIA PhysX simulates it, here with GPU dynamics, and RTX path tracing renders it.
4
+ Isaac Sim is widely used for robot learning and sim-to-real work.
5
+
6
+ Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.
7
+
8
+ THE RESULT MUST BE SIM-READY, READY FOR INTERACTIVE ROBOT LEARNING. It is an environment a robot learns
9
+ in, not a picture of a room. A training pipeline must be able to load it headless, reset it, step it for
10
+ a long time and let robots act on it, with no manual repair. That means:
11
+ - It loads with no network access or external files, and resets to the same state at rest.
12
+ - Everything a robot could touch has collision geometry matching what is visible. An object with no
13
+ collision does not exist for a robot, however good it looks.
14
+ - Rigid things are rigid bodies with plausible mass; free objects can be pushed, and doors, drawers and lids
15
+ move on joints with limits.
16
+ - Soft things are native deformables: cloth, bedding, curtains, rugs, throws, cushions and pillows.
17
+ A rigid body that merely looks soft is wrong.
18
+ - It stays stable under sustained interaction, across many steps without a reset, and steps fast
19
+ enough to train on.
20
+
21
+ Budget: 30 minutes (1,800 seconds) of wall-clock time per scene, starting when
22
+ you receive access to the input files. All inspection, coding, asset creation,
23
+ tool calls, rendering, debugging, and validation must fit within this budget.
24
+ Stop work at the deadline and submit the latest valid, loadable result.
25
+ Do not stop after producing the first runnable scene. Use the available time
26
+ for repeated self-critique and refinement; reserve the final 90 seconds for
27
+ validation and submission.
28
+
29
+ Inputs (in /work/input):
30
+ - video.mp4: the complete source video.
31
+ - scale_anchor.json: dimensions and identity of one object in the room, the scale anchor.
32
+ - anchor_reference.png and anchor_marked.png: images identifying that object.
33
+ The marked rectangle is a projected box envelope, not a segmentation mask.
34
+ - generation_spec.json: this contract in machine-readable form.
35
+ - self_check.py: the public self-check described below.
36
+
37
+ Maximize visual realism and fidelity to the source video. Reproduce the room's
38
+ geometry, proportions, native objects, materials, textures, colors, lighting,
39
+ and visible details as faithfully as possible within the time budget. Use the
40
+ scale anchor to establish metric scale; infer other dimensions and layout from
41
+ the video. The anchor's known real dimensions let you work out every other
42
+ object's size and the distances between objects, for example by measuring them
43
+ against the anchor in frames where both are visible, or by scaling a
44
+ reconstruction so that the anchor has its true size. Place the anchor object
45
+ where it is in the video; the room is not aligned to it alone.
46
+
47
+ Visual objective: MAXIMIZE THE VISUAL QUALITY of the reconstructed room, without giving up any
48
+ sim-ready requirement: a scene that looks right but cannot be simulated has failed.
49
+ Use source-aligned renders to improve what a viewer actually sees: recognizable
50
+ furniture silhouettes, correct proportions and placement, detailed surfaces,
51
+ plausible texture scale, window appearance, soft lighting and small distinctive
52
+ objects. Avoid a room made only of plain boxes when the video shows richer
53
+ shapes and materials. Create local textures or meshes when they improve fidelity;
54
+ do not use a single photograph or video plane as a substitute for 3D geometry.
55
+ Check the entire camera path, including corners and reverse views, rather than
56
+ optimizing one attractive render.
57
+
58
+ Required self-critique and refinement:
59
+ 1. Inspect at least 12 frames distributed across the video. Choose at least
60
+ six representative reference views covering the room and important objects.
61
+ Record their timestamps. Estimate corresponding cameras using ONLY the
62
+ public video and scale anchor, and reuse those cameras across iterations.
63
+ 2. Build an early loadable baseline and render those views. Do not spend the
64
+ whole budget on a first draft or a single attractive camera angle.
65
+ 3. Complete at least THREE render-critique-refine-verify rounds. In each round,
66
+ open side-by-side source/render contact sheets with the image-capable Read
67
+ tool. Identify concrete visible discrepancies, implement the highest-impact
68
+ fixes, then re-render the SAME views and check for regressions. Writing a
69
+ critique without changing and re-rendering the scene is not a completed round.
70
+ 4. Round 1 prioritizes room shape, scale, furniture placement, object dimensions
71
+ and silhouettes. Round 2 prioritizes materials, textures, lighting, colors,
72
+ windows and salient details. Round 3 revisits visual mismatches and verifies
73
+ collision support, movable parts and soft items, and runs the self-check.
74
+ 5. Continue additional rounds while time permits. Do not wait until the deadline
75
+ to save a valid scene. Report the true number of completed rounds.
76
+ 6. After each round, save a valid checkpoint and record the source timestamps,
77
+ render paths, specific discrepancies, actual edits, before/after observations,
78
+ and remaining issues in refinement_log.json. Treat self-assessments as
79
+ diagnostics.
80
+ 7. Keep the best validated complete scene, and restore it if a later change
81
+ regresses. The final output must describe that selected result.
82
+
83
+ You may use any tools, including web search, and the software installed here,
84
+ including Blender (`blender`). Download nothing from the internet (no files,
85
+ models, assets, datasets or packages), and do not look up this room, its video
86
+ or the dataset it comes from. Use no annotations, scans or other data about this
87
+ room beyond your inputs, no previous generated outputs and no other session's
88
+ reconstruction. Do not read authentication files. Do not add a robot.
89
+
90
+ The result must be a genuinely interactive 3D scene:
91
+ - Preserve native object identities, count, dimensions, and spatial relationships.
92
+ - Provide collision geometry consistent with visible surfaces and free spaces.
93
+ - Separate movable objects and articulated parts, with appropriate joints and limits.
94
+ - Build every visible soft item with the simulator's native deformable mechanism, as one object
95
+ entry each, with a label that says plainly what it is, for example "curtain".
96
+ - Preserve openings, container interiors, support surfaces, and access needed
97
+ for interactions. Do not replace the room's own objects with generic stand-ins.
98
+ - Assign physically plausible properties and document estimates for unknowns;
99
+ the video is not measured mass, friction or stiffness.
100
+
101
+ Construction rules (the self-check reports them):
102
+ - The scene must start at rest without interpenetration: no contact deeper than
103
+ 5 mm between any two bodies or between a body and fixed geometry.
104
+ - Articulated parts (drawers, doors, lids) must not intersect their own carcass in
105
+ any position of their range: hollow the carcass or exclude the pair from collision.
106
+ - Door and drawer swing paths must be clear of other furniture.
107
+ - The timestep is your choice.
108
+
109
+ Object labels: each objects.json entry has a "label" that says plainly what the object is
110
+ (for example "double bed", "floor lamp"). Use one entry per physical instance.
111
+
112
+ Interaction points: for articulated furniture give a "handle" point on each moving
113
+ part; for movable graspable objects a "grasp" point; for pourable containers a
114
+ "spout" point; for a visible bed cover, four "material_corners" in boundary order.
115
+
116
+ Self-check: run it on your output after each round and before submitting. It loads
117
+ only your output, simulates 3 s from the authored state and reports contacts deeper
118
+ than 5 mm (with the two bodies involved), the time above 5 mm, peak body speed,
119
+ non-finite state and entries without a label. The scene must stay within these
120
+ limits: at most 0.02 s above 5 mm, peak speed at most 5 m/s, finite state. Fix what it reports; do not edit it.
121
+ /isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check
122
+
123
+ Output format (Isaac Sim, in /work/output):
124
+ - scene.usda (or scene.usd): self-contained USD stage, Z-up, metres, explicit default
125
+ prim, locally referenced assets and textures, and an authored PhysX scene.
126
+ - build_scene.py: reproducible builder using the installed Isaac Sim / pxr APIs.
127
+ - objects.json: {"objects": [...]}, one entry per physical object instance:
128
+ "id": a persistent unique string (keep the scale anchor ID),
129
+ "label": what the object is, in plain words,
130
+ "prim_path": the object's root prim (no single mesh containing all furniture),
131
+ "joint_paths": articulated joint prims, if present,
132
+ "interaction_points": interaction Xform prim paths, named handle, grasp or spout
133
+ (child Xforms of the moving part or object),
134
+ "material_corners": four mesh vertex indices for a deformable bed cover.
135
+ - render_preview.py, at least six views rendered with RTX PathTracing, preview.mp4
136
+ (camera tour) and interaction.mp4 (native physics interaction with an original object).
137
+ - README.md and refinement_log.json, as above.
138
+ Runtime: /isaac-sim/python.sh YOUR_SCRIPT.py runs the bundled Python with Isaac Sim,
139
+ USD, NumPy and Pillow; /usr/local/bin/ffmpeg and blender are available. Use headless=True,
140
+ multi_gpu=False, active_gpu=0, physics_gpu=0. sdk_smoke.py is an API and rendering
141
+ example. Native RTX capture can need several app updates before RGB has shape
142
+ (height, width, 4). Only input and output are host-mounted. Do not access
143
+ authentication.
scenes/isaac/47331524/20261114T061812Z/input/scale_anchor.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "object_id": "scale_anchor",
3
+ "description": "The stool outlined in red in the marked image.",
4
+ "reference_image": "anchor_reference.png",
5
+ "marked_image": "anchor_marked.png",
6
+ "region_xyxy_pixels": [
7
+ 158.55090286972978,
8
+ 89.35109928731454,
9
+ 245.7067610769665,
10
+ 185.8587401803795
11
+ ],
12
+ "region_type": "projected_3d_box_envelope_not_segmentation",
13
+ "size_local_xyz_m": [
14
+ 0.33370427970979016,
15
+ 0.3534851293272629,
16
+ 0.4539529003662027
17
+ ],
18
+ "axis_description": [
19
+ "horizontal depth",
20
+ "horizontal width",
21
+ "vertical height"
22
+ ],
23
+ "measurement_type": "full extents of an oriented box around the object",
24
+ "uncertainty_m": null,
25
+ "uncertainty_note": "No uncertainty is supplied for these dimensions."
26
+ }
scenes/isaac/47331524/20261114T061812Z/input/sdk_smoke.py ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Verify Isaac Sim headless RTX rendering and native rigid-body stepping."""
2
+ from isaacsim import SimulationApp
3
+ app = SimulationApp({'headless':True,'renderer':'PathTracing','samples_per_pixel_per_frame':32,
4
+ 'width':512,'height':384,'multi_gpu':False,'active_gpu':0,'physics_gpu':0,
5
+ 'enable_crashreporter':False,'limit_cpu_threads':8})
6
+ import json
7
+ from pathlib import Path
8
+ import numpy as np
9
+ from PIL import Image
10
+ import omni.usd
11
+ import omni.replicator.core as rep
12
+ rep.orchestrator.set_capture_on_play(False)
13
+ from pxr import UsdGeom,UsdLux,UsdShade,UsdPhysics,Gf,Sdf
14
+ stage=omni.usd.get_context().get_stage()
15
+ UsdGeom.SetStageUpAxis(stage,UsdGeom.Tokens.z);UsdGeom.SetStageMetersPerUnit(stage,1.)
16
+ UsdGeom.Xform.Define(stage,'/World');stage.SetDefaultPrim(stage.GetPrimAtPath('/World'))
17
+ def box(path,pos,scale,color,dynamic=False):
18
+ prim=UsdGeom.Cube.Define(stage,path);prim.CreateSizeAttr(1.)
19
+ xf=UsdGeom.Xformable(prim);xf.AddTranslateOp().Set(Gf.Vec3d(*pos));xf.AddScaleOp().Set(Gf.Vec3f(*scale))
20
+ UsdPhysics.CollisionAPI.Apply(prim.GetPrim())
21
+ if dynamic:
22
+ UsdPhysics.RigidBodyAPI.Apply(prim.GetPrim());UsdPhysics.MassAPI.Apply(prim.GetPrim()).CreateMassAttr(1.)
23
+ mat=UsdShade.Material.Define(stage,path+'_Material');shader=UsdShade.Shader.Define(stage,path+'_Material/Shader');shader.CreateIdAttr('UsdPreviewSurface')
24
+ shader.CreateInput('diffuseColor',Sdf.ValueTypeNames.Color3f).Set(Gf.Vec3f(*color));shader.CreateInput('roughness',Sdf.ValueTypeNames.Float).Set(.28)
25
+ mat.CreateSurfaceOutput().ConnectToSource(shader.ConnectableAPI(),'surface');UsdShade.MaterialBindingAPI.Apply(prim.GetPrim()).Bind(mat)
26
+ return prim
27
+ box('/World/Floor',(0,0,-.05),(5,5,.1),(.6,.6,.6));box('/World/Cube',(0,0,1.),(.3,.3,.3),(.08,.3,.8),True)
28
+ light=UsdLux.DomeLight.Define(stage,'/World/Dome');light.CreateIntensityAttr(500.)
29
+ light=UsdLux.RectLight.Define(stage,'/World/Key');light.CreateIntensityAttr(2500.);light.CreateWidthAttr(3.);light.CreateHeightAttr(3.);UsdGeom.Xformable(light).AddTranslateOp().Set(Gf.Vec3d(0,0,3))
30
+ scene=UsdPhysics.Scene.Define(stage,'/World/PhysicsScene');scene.CreateGravityDirectionAttr(Gf.Vec3f(0,0,-1));scene.CreateGravityMagnitudeAttr(9.81)
31
+ camera=rep.create.camera(position=(2.4,-3,2),look_at=(0,0,.3));product=rep.create.render_product(camera,(512,384));rgb=rep.AnnotatorRegistry.get_annotator('rgb');rgb.attach([product])
32
+ from isaacsim.core.api import World
33
+ world=World(physics_dt=1/60,rendering_dt=1/60,stage_units_in_meters=1.)
34
+ world.reset()
35
+ for _ in range(30):world.step(render=True)
36
+ for attempt in range(30):
37
+ rep.orchestrator.step(rt_subframes=8)
38
+ app.update()
39
+ image=np.asarray(rgb.get_data())
40
+ print('CAPTURE_ATTEMPT',attempt,'SHAPE',image.shape,flush=True)
41
+ if image.ndim==3 and image.shape[0]>0:break
42
+ print('RGB_SHAPE',image.shape,flush=True)
43
+ assert image.ndim==3,'RTX annotator did not return a frame'
44
+ position=UsdGeom.Xformable(stage.GetPrimAtPath('/World/Cube')).ComputeLocalToWorldTransform(0).ExtractTranslation()
45
+ print('CUBE_POSITION',list(position),flush=True)
46
+ out=Path('/work/output');out.mkdir(exist_ok=True)
47
+ Image.fromarray(image[:,:,:3]).save(out/'smoke.png');stage.GetRootLayer().Export(str(out/'sdk_example.usda'))
48
+ (out/'smoke.json').write_text(json.dumps({'renderer':'Isaac Sim RTX PathTracing','shape':list(image.shape),'cube_final_position':list(position),'initial_height_m':1.,'success':True}))
49
+ app.close()
scenes/isaac/47331524/20261114T061812Z/input/self_check.py ADDED
@@ -0,0 +1,228 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Passive physics of a USD scene in Isaac Sim, against the scene requirements (no penetration beyond 5 mm for more than
2
+ 0.02 s, no body faster than 5 m/s, finite state).
3
+
4
+ Runs inside the Isaac Sim container. Reads --scene (default /work/scene, read-only) and the optional
5
+ --out/request.json (object ids to drop); writes --out/physics.json (default /work/physics). Without a request it is
6
+ the self-check: no drops, plus a check that every object has a non-empty label.
7
+
8
+ - Speed: a 3 s passive rollout with no rendering and no state reads, timed around world.step only
9
+ (simulated seconds / wall seconds).
10
+ - Passive rollout: 3 s from the authored state. Penetration comes from the PhysX tensor contact view
11
+ (GPU contact reports return nothing under GPU dynamics); speed and poses from the RigidPrim view.
12
+ - Drops: the requested objects with free (unjointed) dynamic bodies, raised 0.05 m
13
+ and released for 2 s, each from a fresh reset.
14
+ The limits are the same as for MuJoCo scenes.
15
+ """
16
+ from isaacsim import SimulationApp
17
+ app=SimulationApp({'headless':True,'multi_gpu':False,'active_gpu':0,'physics_gpu':0,
18
+ 'enable_crashreporter':False,'limit_cpu_threads':8})
19
+ import argparse
20
+ import json
21
+ from pathlib import Path
22
+ import time
23
+ import carb
24
+ import numpy as np
25
+ import omni.usd
26
+ from pxr import UsdPhysics,PhysxSchema
27
+
28
+ LIMIT_M,LIMIT_S,SPEED=.005,.02,5.
29
+ parser=argparse.ArgumentParser();parser.add_argument('--scene',default='/work/scene');parser.add_argument('--out',default='/work/physics')
30
+ args,_=parser.parse_known_args()
31
+ out=Path(args.out);out.mkdir(parents=True,exist_ok=True)
32
+ request=json.loads((out/'request.json').read_text()) if (out/'request.json').is_file() else {'object_ids':[]}
33
+ result={'status':'failed','engine':'Isaac Sim PhysX (GPU dynamics)','no_vlm':True,'provisional':True,
34
+ 'thresholds':{'penetration_m':LIMIT_M,'penetration_duration_s':LIMIT_S,'linear_speed_mps':SPEED},
35
+ 'thresholds_calibrated_across_engines':False}
36
+
37
+
38
+ def contact_filters(stage,paths):
39
+ """GPU contact views filter on rigid bodies or standalone colliders only, so a collider inside a
40
+ rigid body is replaced by that body (unsupported filters are dropped with only a warning)."""
41
+ out=[]
42
+ for path in paths:
43
+ if any('deformable' in str(a).lower() for a in stage.GetPrimAtPath(path).GetAppliedSchemas()):
44
+ continue # deformable meshes cannot be contact filters; their contacts are not measured
45
+ prim=stage.GetPrimAtPath(path);owner=None
46
+ while prim and not prim.IsPseudoRoot():
47
+ if prim.HasAPI(UsdPhysics.RigidBodyAPI):owner=str(prim.GetPath())
48
+ prim=prim.GetParent()
49
+ target=owner or path
50
+ if target not in out:out.append(target)
51
+ return out
52
+
53
+
54
+ def array(value):
55
+ return np.asarray(value.cpu() if hasattr(value,'cpu') else value,dtype=float)
56
+
57
+
58
+ try:
59
+ root=Path(args.scene)
60
+ scene=next(root/n for n in ['scene.usda','scene.usd'] if (root/n).is_file())
61
+ omni.usd.get_context().open_stage(str(scene));stage=omni.usd.get_context().get_stage()
62
+ for _ in range(8):app.update()
63
+
64
+ def dynamic(prim):
65
+ api=UsdPhysics.RigidBodyAPI(prim)
66
+ if api.GetRigidBodyEnabledAttr().Get() not in (None,True) or api.GetKinematicEnabledAttr().Get():return False
67
+ parent=prim.GetParent()
68
+ while parent and not parent.IsPseudoRoot():
69
+ if parent.HasAPI(UsdPhysics.RigidBodyAPI):return False
70
+ parent=parent.GetParent()
71
+ return True
72
+ rigid=[p for p in stage.Traverse() if p.HasAPI(UsdPhysics.RigidBodyAPI)]
73
+ bodies=[p for p in rigid if dynamic(p)]
74
+ paths=[str(p.GetPath()) for p in bodies]
75
+ colliders=[str(p.GetPath()) for p in stage.Traverse() if p.HasAPI(UsdPhysics.CollisionAPI)]
76
+ jointed=set()
77
+ for prim in stage.Traverse():
78
+ if prim.IsA(UsdPhysics.Joint):
79
+ joint=UsdPhysics.Joint(prim)
80
+ for rel in (joint.GetBody0Rel(),joint.GetBody1Rel()):
81
+ jointed.update(str(t) for t in rel.GetTargets())
82
+ for prim in bodies:PhysxSchema.PhysxContactReportAPI.Apply(prim).CreateThresholdAttr().Set(0.)
83
+ scenes=[p for p in stage.Traverse() if p.IsA(UsdPhysics.Scene)]
84
+ authored=PhysxSchema.PhysxSceneAPI(scenes[0]).GetTimeStepsPerSecondAttr().Get() if scenes else None
85
+ rate=float(authored or 60)
86
+ result.update(dynamic_bodies=len(paths),colliders=len(colliders),
87
+ excluded_bodies=[str(p.GetPath()) for p in rigid if not dynamic(p)],
88
+ timestep_s=1/rate,timestep_source='authored PhysxSceneAPI timeStepsPerSecond' if authored else 'default 60 Hz')
89
+ if not paths:raise ValueError('No dynamic rigid bodies to simulate')
90
+ filters=contact_filters(stage,colliders)
91
+ result['contact_filters']=len(filters)
92
+
93
+ carb.settings.get_settings().set_bool('/physics/disableContactProcessing',False)
94
+ from isaacsim.core.api import World
95
+ from isaacsim.core.prims import RigidPrim
96
+ from isaacsim.core.simulation_manager import SimulationManager
97
+ world=World(physics_dt=1/rate,rendering_dt=1/rate,stage_units_in_meters=1.,
98
+ physics_prim_path=str(scenes[0].GetPath()) if scenes else '/World/AddedPhysicsScene',
99
+ set_defaults=False,backend='torch',device='cuda:0')
100
+ context=world.get_physics_context()
101
+ context.enable_gpu_dynamics(True);context.set_broadphase_type('GPU')
102
+ view=RigidPrim(paths,name='scene_bodies');world.scene.add(view)
103
+ steps=lambda seconds:int(np.ceil(seconds*rate))
104
+
105
+ def measure_rtf(dt,repeats=3):
106
+ """Simulated seconds / wall seconds of world.step(render=False) for an unread passive rollout from the authored
107
+ state, median over `repeats` after 10 warm-up steps."""
108
+ world.set_simulation_dt(physics_dt=dt,rendering_dt=dt)
109
+ world.reset()
110
+ for _ in range(10):world.step(render=False)
111
+ count=int(np.ceil(3/dt));samples=[];finite=True
112
+ for _ in range(repeats):
113
+ world.reset();start=time.perf_counter()
114
+ for _ in range(count):world.step(render=False)
115
+ samples.append(count*dt/(time.perf_counter()-start))
116
+ positions,_=view.get_world_poses()
117
+ finite&=bool(np.isfinite(array(positions)).all() and np.isfinite(array(view.get_velocities())).all())
118
+ return {'rtf':float(np.median(samples)),'rtf_samples':samples,'timestep_s':dt,'simulated_seconds':count*dt,
119
+ 'steps':count,'repeats':repeats,'finite':finite,
120
+ 'definition':'simulated seconds / wall seconds of world.step(render=False) only, median of repeats'}
121
+
122
+ result['realtime']={**measure_rtf(1/rate),'scope':'3 s passive rollout from the authored state, authored timestep'}
123
+ print('PHYSICS_STAGE realtime',result['realtime']['rtf'],flush=True)
124
+
125
+ def reset():
126
+ world.reset()
127
+ contacts=SimulationManager.get_physics_sim_view().create_rigid_contact_view(
128
+ paths,filter_patterns=[filters]*len(paths),max_contact_data_count=200000)
129
+ if contacts is None:raise RuntimeError('PhysX rigid contact view could not be created')
130
+ return contacts
131
+
132
+ def rollout(seconds,contacts,motion=None):
133
+ """One rollout; `motion` names the file the logged poses are written to."""
134
+ report={'finite':True,'peak_penetration_m':0.,'penetration_time_s':0.,'peak_linear_speed_mps':0.}
135
+ every=max(1,round(rate/10));frames=[]
136
+ for step in range(steps(seconds)):
137
+ world.step(render=False)
138
+ positions,orientations=view.get_world_poses();velocities=array(view.get_velocities())
139
+ if step%every==0:frames.append(((step+1)/rate,array(positions),array(orientations)))
140
+ report['finite']&=bool(np.isfinite(array(positions)).all() and np.isfinite(velocities).all())
141
+ _,_,_,separation,counts,starts=contacts.get_contact_data(dt=1/rate)
142
+ total=int(array(counts).sum())
143
+ if 'initial_pairs' not in report:
144
+ # Pairs deeper than the limit after the first step, named for the self-check.
145
+ pairs=[];sep=array(separation);cnt=array(counts).astype(int);beg=array(starts).astype(int)
146
+ for i,j in zip(*np.nonzero(cnt)):
147
+ depth=-float(sep[beg[i,j]:beg[i,j]+cnt[i,j]].min())
148
+ if depth>LIMIT_M:pairs.append({'depth_m':depth,'bodies':[paths[i],filters[j]]})
149
+ report['initial_pairs']=sorted(pairs,key=lambda p:-p['depth_m'])[:20]
150
+ deepest=max(0.,-float(array(separation)[:total].min())) if total else 0.
151
+ report['peak_penetration_m']=max(report['peak_penetration_m'],deepest)
152
+ if deepest>LIMIT_M:report['penetration_time_s']+=1/rate
153
+ report['peak_linear_speed_mps']=max(report['peak_linear_speed_mps'],float(np.linalg.norm(velocities[:,:3],axis=1).max()))
154
+ if not report['finite']:break
155
+ report['valid']=bool(report['finite'] and report['penetration_time_s']<=LIMIT_S and report['peak_linear_speed_mps']<=SPEED)
156
+ report['thresholds']=result['thresholds']
157
+ if motion and request.get('record'):
158
+ (out/'motion').mkdir(exist_ok=True)
159
+ np.savez_compressed(out/'motion'/motion,paths=np.array(paths),t=np.array([f[0] for f in frames]),
160
+ positions=np.stack([f[1] for f in frames]),orientations=np.stack([f[2] for f in frames]))
161
+ report['motion']=f'motion/{motion}.npz'
162
+ return report
163
+
164
+ contacts=reset()
165
+ start_positions=array(view.get_world_poses()[0])
166
+ masses=array(view.get_masses())
167
+ direction,magnitude=context.get_gravity()
168
+ gravity=np.asarray(direction,dtype=float)*float(magnitude)
169
+ passive=rollout(3,contacts,'passive')
170
+ print('PHYSICS_STAGE passive',passive,flush=True)
171
+ end_positions=array(view.get_world_poses()[0])
172
+ moved=np.linalg.norm(end_positions-start_positions,axis=1)
173
+
174
+ manifest=json.loads((root/'objects.json').read_text())
175
+ manifest=manifest['objects'] if isinstance(manifest,dict) else manifest
176
+ result['manifest_errors']=[f"{o.get('id')}: label is missing or empty" for o in manifest if not str(o.get('label') or '').strip()]
177
+ result['manifest_errors']+=[f"{o.get('id')}: prim_path does not resolve" for o in manifest if not stage.GetPrimAtPath(o.get('prim_path','') or '/')]
178
+ def owned(obj):
179
+ prefix=str(obj.get('prim_path','')).rstrip('/')
180
+ return [i for i,p in enumerate(paths) if prefix and (p==prefix or p.startswith(prefix+'/'))]
181
+ drift=[{'id':o['id'],'dynamic_bodies':len(owned(o)),'center_drift_m':float(moved[owned(o)].max()),'mass_kg':float(masses[owned(o)].sum())}
182
+ for o in manifest if owned(o)]
183
+
184
+ requested=set(request['object_ids'])
185
+ eligible=[o for o in manifest if o['id'] in requested]
186
+ drops=[]
187
+ for obj in eligible:
188
+ free=[i for i in owned(obj) if paths[i] not in jointed]
189
+ if not free:continue
190
+ contacts=reset()
191
+ positions,orientations=view.get_world_poses()
192
+ raised=positions.clone() if hasattr(positions,'clone') else positions.copy()
193
+ raised[free,2]+=.05
194
+ view.set_world_poses(positions=raised,orientations=orientations)
195
+ applied=float((array(view.get_world_poses()[0])[free,2]-array(positions)[free,2]).min())
196
+ drops.append({'id':obj['id'],'input':'raise free bodies 0.05 m and release','bodies':[paths[i] for i in free],
197
+ 'lift_applied_m':applied,**rollout(2,contacts,f'drop_{len(drops)}')})
198
+ print('PHYSICS_STAGE drop',obj['id'],applied,drops[-1]['peak_linear_speed_mps'],flush=True)
199
+
200
+ # The common 2 ms step last, so the rollouts above run at the authored step.
201
+ result['realtime_common_step']={**measure_rtf(.002),'scope':'3 s passive rollout from the authored state at the common 2 ms step'}
202
+ print('PHYSICS_STAGE realtime_common_step',result['realtime_common_step']['rtf'],flush=True)
203
+ gravity_ok=bool(np.allclose(gravity,[0,0,-9.81],atol=.1))
204
+ mass_ok=bool(np.isfinite(masses).all() and (masses>=0).all())
205
+ result.update(status='completed',passive=passive,gravity_m_s2=gravity.tolist(),gravity_valid=gravity_ok,
206
+ mass_valid=mass_ok,center_drift=drift,drop_tests=drops,eligible_native_objects=len(eligible),
207
+ largest_body_displacements=[{'prim':paths[i],'displacement_m':float(moved[i])} for i in np.argsort(-moved)[:5]],
208
+ valid=bool(passive['valid'] and gravity_ok and mass_ok),
209
+ drop_scope='the requested objects with free (unjointed) dynamic bodies, in manifest order',
210
+ contact_scope=('contacts between each dynamic body and every collider (static, kinematic and dynamic; colliders '
211
+ 'inside a rigid body are filtered by that body, as GPU contact views require); '
212
+ 'static-static pairs are not simulated, as in MuJoCo world-body geoms'),
213
+ limitations=['the limits are the MuJoCo ones, not calibrated for PhysX',
214
+ 'kinematic, disabled and nested rigid bodies are not traced (see excluded_bodies)',
215
+ 'drift is per dynamic body under the object prim, not the object bounding box',
216
+ 'contacts with deformable bodies are not measured (GPU contact views cannot filter them)'])
217
+ except Exception as error:
218
+ result['error']=f'{type(error).__name__}: {error}'
219
+ finally:
220
+ (out/'physics.json').write_text(json.dumps(result,indent=2)+'\n')
221
+ if 'passive' in result:
222
+ passive=result['passive']
223
+ print('PASSIVE CHECK','PASSED' if passive['valid'] and not result.get('manifest_errors') else 'FAILED',
224
+ f"- {len(passive.get('initial_pairs',[]))} contact pairs > 5 mm, {passive['penetration_time_s']:.3f} s above 5 mm,"
225
+ f" peak speed {passive['peak_linear_speed_mps']:.2f} m/s, {len(result.get('manifest_errors',[]))} manifest errors;"
226
+ f" details in {out/'physics.json'}",flush=True)
227
+ else:print('CHECK FAILED:',result.get('error'),flush=True)
228
+ app.close()
scenes/isaac/47331524/20261114T061812Z/input/video.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9ae6725724ceeffa029290fbb935e9b3e7898120b6b9b3b5af5383584c718dd7
3
+ size 5598485
scenes/isaac/47331524/20261114T061812Z/run.json ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "track": "scenes",
3
+ "engine": "isaac",
4
+ "item": "47331524",
5
+ "run_id": "20261114T061812Z",
6
+ "campaign": "jiarui-2-codebuddy",
7
+ "cli": "codebuddy",
8
+ "model": "Hunyuan 4",
9
+ "model_id": "hy4-preview",
10
+ "model_family": "hy4-preview",
11
+ "model_reasoning_effort": "default",
12
+ "budget_seconds": 1800,
13
+ "told_budget_seconds": 1800,
14
+ "stage": 1,
15
+ "continued_from": null,
16
+ "contract": {
17
+ "name": "scenes-isaac",
18
+ "sha256": "f1ab7ea88842502d37176b41d9f3bec64a5d9fac292ca3f16caf08da3b5db96d",
19
+ "track": "scenes",
20
+ "engine": "isaac",
21
+ "engine_version": "6.2.0",
22
+ "image": "simready-isaac-generation:latest",
23
+ "blender_version": "5.2.3",
24
+ "inputs": [
25
+ "video.mp4",
26
+ "scale_anchor.json",
27
+ "anchor_reference.png",
28
+ "anchor_marked.png",
29
+ "prompt.txt",
30
+ "generation_spec.json",
31
+ "self_check.py",
32
+ "sdk_smoke.py"
33
+ ],
34
+ "withheld": [
35
+ "groundtruth/ (objects, cameras, task catalogs, object dynamics, interaction test) and its source tag",
36
+ "provenance/ (source annotation, anchor selection)",
37
+ "the frozen protocol of the track and engine (data/benchmark/frozen/<track>/<engine>.json)",
38
+ "other runs"
39
+ ],
40
+ "outputs": [
41
+ "scene.usda",
42
+ "build_scene.py",
43
+ "objects.json",
44
+ "render_preview.py",
45
+ "preview.mp4",
46
+ "interaction.mp4",
47
+ "README.md",
48
+ "refinement_log.json"
49
+ ],
50
+ "self_check": "/isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check",
51
+ "budget": {
52
+ "told": "the first of the session's cumulative budgets, in wall-clock seconds from container start",
53
+ "resume": "each later budget resumes the same conversation for the rest of its time; each budget is its own run",
54
+ "enforcement": "the runner kills the container at the deadline; the output as it stands is the result"
55
+ },
56
+ "refinement": {
57
+ "minimum_completed_rounds": 3,
58
+ "minimum_reference_views": 6,
59
+ "minimum_inspected_video_frames": 12,
60
+ "final_validation_seconds": 90
61
+ },
62
+ "isolation": {
63
+ "container": "Docker: read-only input (and tools), writable output and scratch, temporary credentials; no repository, benchmark ground truth or evaluator files",
64
+ "network": "an internal network whose only exit is an HTTPS CONNECT proxy (runtime/egress_proxy.py)",
65
+ "egress_hosts": [
66
+ "api.anthropic.com",
67
+ "platform.claude.com",
68
+ "chatgpt.com",
69
+ "auth.openai.com",
70
+ "opencode.ai",
71
+ "mcp.exa.ai",
72
+ "search.parallel.ai"
73
+ ]
74
+ },
75
+ "protocol": {
76
+ "file": "data/benchmark/frozen/scenes/isaac.json",
77
+ "sha256": "df0fe5bb1c796474ad13cdba81b660a2946aa24297cb289f9581cfde7c035bb7"
78
+ }
79
+ },
80
+ "engine_version": "6.2.0",
81
+ "image": "simready-isaac-generation:latest",
82
+ "image_id": "sha256:8d9f9c934457aa3c8e91e8099e35db27b50db287d0eb76a647052c310e754d87",
83
+ "required_refinement_rounds": 3,
84
+ "gpu_uuid": "GPU-ab037b41-847d-64be-b2b8-ba682a201943",
85
+ "egress_mode": "internal_network",
86
+ "input_snapshot": "input",
87
+ "tools": null,
88
+ "evaluator_protocol": "evaluator_protocol.json",
89
+ "output": "output",
90
+ "log": "generation.jsonl",
91
+ "status": "budget_expired",
92
+ "agent_cli": {
93
+ "cli": "codebuddy",
94
+ "name": "CodeBuddy Code",
95
+ "binary": "a pinned codebuddy-code package and its node, mounted read-only (SIMREADY_CODEBUDDY_PREFIX, default the host's codebuddy install)",
96
+ "login": "the CodeBuddy login file, given to the CLI as a snapshot without its refresh token (the session cannot rotate it)",
97
+ "billing": "CodeBuddy account (the CLI reports no price)",
98
+ "tools": "every tool but WebFetch and the server-side media generators ImageGen, ImageEdit, VideoGen (--disallowedTools); WebSearch runs on CodeBuddy's side",
99
+ "hosts": [
100
+ "[redacted-host]",
101
+ "wb.[redacted]buddy.com"
102
+ ],
103
+ "version": "2.182.0",
104
+ "model": "hy4-preview",
105
+ "model_family": "hy4-preview"
106
+ },
107
+ "started_at": "2026-11-14T06:18:13.911784+00:00",
108
+ "model_answered": true,
109
+ "exit_code": 137,
110
+ "credential_refresh_persistence": "snapshot_without_refresh_token",
111
+ "finished_at": "2026-11-14T06:48:15.916730+00:00",
112
+ "cli_session_id": "01a1d8b4-e556-764b-a611-3a2b8c2daba1",
113
+ "api_retries": 0,
114
+ "rate_limit_events": 0,
115
+ "stage_seconds": 1802.062,
116
+ "elapsed_seconds": 1802.062,
117
+ "cost": {
118
+ "status": "partial",
119
+ "source": "the CLI's running usage totals (usage_update events)",
120
+ "usd": null,
121
+ "usd_basis": "not reported (CodeBuddy account billing)",
122
+ "credits": 0,
123
+ "credit_status": "known",
124
+ "turns": null,
125
+ "api_seconds": null,
126
+ "tokens": {
127
+ "input": 63269,
128
+ "cache_read": 552320,
129
+ "cache_write": 0,
130
+ "output": 27273,
131
+ "reasoning": null,
132
+ "total": 642862
133
+ },
134
+ "wall_seconds": 1802.062,
135
+ "budget_seconds": 1800,
136
+ "stopped_at_budget": true
137
+ },
138
+ "scene_present": false
139
+ }
scenes/isaac/47331524/20261114T061812Z/submitted_prompt.txt ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ About the simulator. NVIDIA Isaac Sim is a robotics simulator built on Omniverse. You describe a scene
2
+ as an OpenUSD stage: prims carrying physics schemas for rigid bodies, colliders, joints and articulations, and
3
+ deformable bodies and surfaces. NVIDIA PhysX simulates it, here with GPU dynamics, and RTX path tracing renders it.
4
+ Isaac Sim is widely used for robot learning and sim-to-real work.
5
+
6
+ Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.
7
+
8
+ THE RESULT MUST BE SIM-READY, READY FOR INTERACTIVE ROBOT LEARNING. It is an environment a robot learns
9
+ in, not a picture of a room. A training pipeline must be able to load it headless, reset it, step it for
10
+ a long time and let robots act on it, with no manual repair. That means:
11
+ - It loads with no network access or external files, and resets to the same state at rest.
12
+ - Everything a robot could touch has collision geometry matching what is visible. An object with no
13
+ collision does not exist for a robot, however good it looks.
14
+ - Rigid things are rigid bodies with plausible mass; free objects can be pushed, and doors, drawers and lids
15
+ move on joints with limits.
16
+ - Soft things are native deformables: cloth, bedding, curtains, rugs, throws, cushions and pillows.
17
+ A rigid body that merely looks soft is wrong.
18
+ - It stays stable under sustained interaction, across many steps without a reset, and steps fast
19
+ enough to train on.
20
+
21
+ Budget: 30 minutes (1,800 seconds) of wall-clock time per scene, starting when
22
+ you receive access to the input files. All inspection, coding, asset creation,
23
+ tool calls, rendering, debugging, and validation must fit within this budget.
24
+ Stop work at the deadline and submit the latest valid, loadable result.
25
+ Do not stop after producing the first runnable scene. Use the available time
26
+ for repeated self-critique and refinement; reserve the final 90 seconds for
27
+ validation and submission.
28
+
29
+ Inputs (in /work/input):
30
+ - video.mp4: the complete source video.
31
+ - scale_anchor.json: dimensions and identity of one object in the room, the scale anchor.
32
+ - anchor_reference.png and anchor_marked.png: images identifying that object.
33
+ The marked rectangle is a projected box envelope, not a segmentation mask.
34
+ - generation_spec.json: this contract in machine-readable form.
35
+ - self_check.py: the public self-check described below.
36
+
37
+ Maximize visual realism and fidelity to the source video. Reproduce the room's
38
+ geometry, proportions, native objects, materials, textures, colors, lighting,
39
+ and visible details as faithfully as possible within the time budget. Use the
40
+ scale anchor to establish metric scale; infer other dimensions and layout from
41
+ the video. The anchor's known real dimensions let you work out every other
42
+ object's size and the distances between objects, for example by measuring them
43
+ against the anchor in frames where both are visible, or by scaling a
44
+ reconstruction so that the anchor has its true size. Place the anchor object
45
+ where it is in the video; the room is not aligned to it alone.
46
+
47
+ Visual objective: MAXIMIZE THE VISUAL QUALITY of the reconstructed room, without giving up any
48
+ sim-ready requirement: a scene that looks right but cannot be simulated has failed.
49
+ Use source-aligned renders to improve what a viewer actually sees: recognizable
50
+ furniture silhouettes, correct proportions and placement, detailed surfaces,
51
+ plausible texture scale, window appearance, soft lighting and small distinctive
52
+ objects. Avoid a room made only of plain boxes when the video shows richer
53
+ shapes and materials. Create local textures or meshes when they improve fidelity;
54
+ do not use a single photograph or video plane as a substitute for 3D geometry.
55
+ Check the entire camera path, including corners and reverse views, rather than
56
+ optimizing one attractive render.
57
+
58
+ Required self-critique and refinement:
59
+ 1. Inspect at least 12 frames distributed across the video. Choose at least
60
+ six representative reference views covering the room and important objects.
61
+ Record their timestamps. Estimate corresponding cameras using ONLY the
62
+ public video and scale anchor, and reuse those cameras across iterations.
63
+ 2. Build an early loadable baseline and render those views. Do not spend the
64
+ whole budget on a first draft or a single attractive camera angle.
65
+ 3. Complete at least THREE render-critique-refine-verify rounds. In each round,
66
+ open side-by-side source/render contact sheets with the image-capable Read
67
+ tool. Identify concrete visible discrepancies, implement the highest-impact
68
+ fixes, then re-render the SAME views and check for regressions. Writing a
69
+ critique without changing and re-rendering the scene is not a completed round.
70
+ 4. Round 1 prioritizes room shape, scale, furniture placement, object dimensions
71
+ and silhouettes. Round 2 prioritizes materials, textures, lighting, colors,
72
+ windows and salient details. Round 3 revisits visual mismatches and verifies
73
+ collision support, movable parts and soft items, and runs the self-check.
74
+ 5. Continue additional rounds while time permits. Do not wait until the deadline
75
+ to save a valid scene. Report the true number of completed rounds.
76
+ 6. After each round, save a valid checkpoint and record the source timestamps,
77
+ render paths, specific discrepancies, actual edits, before/after observations,
78
+ and remaining issues in refinement_log.json. Treat self-assessments as
79
+ diagnostics.
80
+ 7. Keep the best validated complete scene, and restore it if a later change
81
+ regresses. The final output must describe that selected result.
82
+
83
+ You may use any tools, including web search, and the software installed here,
84
+ including Blender (`blender`). Download nothing from the internet (no files,
85
+ models, assets, datasets or packages), and do not look up this room, its video
86
+ or the dataset it comes from. Use no annotations, scans or other data about this
87
+ room beyond your inputs, no previous generated outputs and no other session's
88
+ reconstruction. Do not read authentication files. Do not add a robot.
89
+
90
+ The result must be a genuinely interactive 3D scene:
91
+ - Preserve native object identities, count, dimensions, and spatial relationships.
92
+ - Provide collision geometry consistent with visible surfaces and free spaces.
93
+ - Separate movable objects and articulated parts, with appropriate joints and limits.
94
+ - Build every visible soft item with the simulator's native deformable mechanism, as one object
95
+ entry each, with a label that says plainly what it is, for example "curtain".
96
+ - Preserve openings, container interiors, support surfaces, and access needed
97
+ for interactions. Do not replace the room's own objects with generic stand-ins.
98
+ - Assign physically plausible properties and document estimates for unknowns;
99
+ the video is not measured mass, friction or stiffness.
100
+
101
+ Construction rules (the self-check reports them):
102
+ - The scene must start at rest without interpenetration: no contact deeper than
103
+ 5 mm between any two bodies or between a body and fixed geometry.
104
+ - Articulated parts (drawers, doors, lids) must not intersect their own carcass in
105
+ any position of their range: hollow the carcass or exclude the pair from collision.
106
+ - Door and drawer swing paths must be clear of other furniture.
107
+ - The timestep is your choice.
108
+
109
+ Object labels: each objects.json entry has a "label" that says plainly what the object is
110
+ (for example "double bed", "floor lamp"). Use one entry per physical instance.
111
+
112
+ Interaction points: for articulated furniture give a "handle" point on each moving
113
+ part; for movable graspable objects a "grasp" point; for pourable containers a
114
+ "spout" point; for a visible bed cover, four "material_corners" in boundary order.
115
+
116
+ Self-check: run it on your output after each round and before submitting. It loads
117
+ only your output, simulates 3 s from the authored state and reports contacts deeper
118
+ than 5 mm (with the two bodies involved), the time above 5 mm, peak body speed,
119
+ non-finite state and entries without a label. The scene must stay within these
120
+ limits: at most 0.02 s above 5 mm, peak speed at most 5 m/s, finite state. Fix what it reports; do not edit it.
121
+ /isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check
122
+
123
+ Output format (Isaac Sim, in /work/output):
124
+ - scene.usda (or scene.usd): self-contained USD stage, Z-up, metres, explicit default
125
+ prim, locally referenced assets and textures, and an authored PhysX scene.
126
+ - build_scene.py: reproducible builder using the installed Isaac Sim / pxr APIs.
127
+ - objects.json: {"objects": [...]}, one entry per physical object instance:
128
+ "id": a persistent unique string (keep the scale anchor ID),
129
+ "label": what the object is, in plain words,
130
+ "prim_path": the object's root prim (no single mesh containing all furniture),
131
+ "joint_paths": articulated joint prims, if present,
132
+ "interaction_points": interaction Xform prim paths, named handle, grasp or spout
133
+ (child Xforms of the moving part or object),
134
+ "material_corners": four mesh vertex indices for a deformable bed cover.
135
+ - render_preview.py, at least six views rendered with RTX PathTracing, preview.mp4
136
+ (camera tour) and interaction.mp4 (native physics interaction with an original object).
137
+ - README.md and refinement_log.json, as above.
138
+ Runtime: /isaac-sim/python.sh YOUR_SCRIPT.py runs the bundled Python with Isaac Sim,
139
+ USD, NumPy and Pillow; /usr/local/bin/ffmpeg and blender are available. Use headless=True,
140
+ multi_gpu=False, active_gpu=0, physics_gpu=0. sdk_smoke.py is an API and rendering
141
+ example. Native RTX capture can need several app updates before RGB has shape
142
+ (height, width, 4). Only input and output are host-mounted. Do not access
143
+ authentication.
scenes/isaac/47331524/20261114T081855Z/evaluator_protocol.json ADDED
The diff for this file is too large to render. See raw diff
 
scenes/isaac/47331524/20261114T081855Z/generation.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
scenes/isaac/47331524/20261114T081855Z/input/anchor_marked.png ADDED

Git LFS Details

  • SHA256: 2ea99bbde1d2d77e45fc55e41276f0edb0d74cebe5e7a0dd8d09c83be372847a
  • Pointer size: 130 Bytes
  • Size of remote file: 74 kB
scenes/isaac/47331524/20261114T081855Z/input/anchor_reference.png ADDED

Git LFS Details

  • SHA256: fed846168acd61f0d575c69a7d522468bbb28d5e57a326f920055dcfd6224f05
  • Pointer size: 130 Bytes
  • Size of remote file: 74.6 kB
scenes/isaac/47331524/20261114T081855Z/input/generation_spec.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "contract": "scenes-isaac",
3
+ "engine": "Isaac Sim",
4
+ "task": "Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.",
5
+ "inputs": [
6
+ "video.mp4",
7
+ "scale_anchor.json",
8
+ "anchor_reference.png",
9
+ "anchor_marked.png",
10
+ "self_check.py"
11
+ ],
12
+ "prompt": "prompt.txt",
13
+ "units": "metres, kilograms, seconds",
14
+ "up_axis": "+Z",
15
+ "entrypoint": "scene.usda",
16
+ "budget": {
17
+ "wall_clock_seconds": 900,
18
+ "scope": "per_scene",
19
+ "starts_at": "agent_input_access",
20
+ "includes_all_tool_and_validation_time": true,
21
+ "deadline_action": "stop and submit latest valid loadable output",
22
+ "enforcement": "Docker generation runner terminates the container at the wall-clock deadline"
23
+ },
24
+ "requirement": "Sim-ready: ready for interactive robot learning, with no manual repair.",
25
+ "sim_ready": {
26
+ "loads_headless_offline": true,
27
+ "resets_to_rest": true,
28
+ "collision_matches_visible_geometry": true,
29
+ "rigid_things": "rigid bodies; free objects movable, articulated parts on limited joints",
30
+ "soft_things": [
31
+ "cloth",
32
+ "bedding",
33
+ "curtains",
34
+ "rugs",
35
+ "throws",
36
+ "cushions",
37
+ "pillows"
38
+ ],
39
+ "soft_mechanism": "native deformable",
40
+ "stable_under_sustained_interaction": true
41
+ },
42
+ "visual_objective": "Maximize visual quality, realism and fidelity to the source video, keeping every sim-ready requirement.",
43
+ "refinement": {
44
+ "minimum_completed_rounds": 3,
45
+ "minimum_reference_views": 6,
46
+ "minimum_inspected_video_frames": 12,
47
+ "final_validation_seconds": 90,
48
+ "outside_data_allowed": false
49
+ },
50
+ "label": "free text saying plainly what the object is; one entry per physical instance",
51
+ "interaction_points": [
52
+ "handle",
53
+ "grasp",
54
+ "spout",
55
+ "material_corners"
56
+ ],
57
+ "construction_rules": {
58
+ "max_initial_penetration_m": 0.005,
59
+ "max_penetration_duration_s": 0.02,
60
+ "max_passive_speed_mps": 5.0,
61
+ "articulated_parts_clear_of_carcass": true,
62
+ "swing_paths_clear": true,
63
+ "timestep": "agent's choice"
64
+ },
65
+ "self_check": "/isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check"
66
+ }
scenes/isaac/47331524/20261114T081855Z/input/prompt.txt ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ About the simulator. NVIDIA Isaac Sim is a robotics simulator built on Omniverse. You describe a scene
2
+ as an OpenUSD stage: prims carrying physics schemas for rigid bodies, colliders, joints and articulations, and
3
+ deformable bodies and surfaces. NVIDIA PhysX simulates it, here with GPU dynamics, and RTX path tracing renders it.
4
+ Isaac Sim is widely used for robot learning and sim-to-real work.
5
+
6
+ Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.
7
+
8
+ THE RESULT MUST BE SIM-READY, READY FOR INTERACTIVE ROBOT LEARNING. It is an environment a robot learns
9
+ in, not a picture of a room. A training pipeline must be able to load it headless, reset it, step it for
10
+ a long time and let robots act on it, with no manual repair. That means:
11
+ - It loads with no network access or external files, and resets to the same state at rest.
12
+ - Everything a robot could touch has collision geometry matching what is visible. An object with no
13
+ collision does not exist for a robot, however good it looks.
14
+ - Rigid things are rigid bodies with plausible mass; free objects can be pushed, and doors, drawers and lids
15
+ move on joints with limits.
16
+ - Soft things are native deformables: cloth, bedding, curtains, rugs, throws, cushions and pillows.
17
+ A rigid body that merely looks soft is wrong.
18
+ - It stays stable under sustained interaction, across many steps without a reset, and steps fast
19
+ enough to train on.
20
+
21
+ Budget: 15 minutes (900 seconds) of wall-clock time per scene, starting when
22
+ you receive access to the input files. All inspection, coding, asset creation,
23
+ tool calls, rendering, debugging, and validation must fit within this budget.
24
+ Stop work at the deadline and submit the latest valid, loadable result.
25
+ Do not stop after producing the first runnable scene. Use the available time
26
+ for repeated self-critique and refinement; reserve the final 90 seconds for
27
+ validation and submission.
28
+
29
+ Inputs (in /work/input):
30
+ - video.mp4: the complete source video.
31
+ - scale_anchor.json: dimensions and identity of one object in the room, the scale anchor.
32
+ - anchor_reference.png and anchor_marked.png: images identifying that object.
33
+ The marked rectangle is a projected box envelope, not a segmentation mask.
34
+ - generation_spec.json: this contract in machine-readable form.
35
+ - self_check.py: the public self-check described below.
36
+
37
+ Maximize visual realism and fidelity to the source video. Reproduce the room's
38
+ geometry, proportions, native objects, materials, textures, colors, lighting,
39
+ and visible details as faithfully as possible within the time budget. Use the
40
+ scale anchor to establish metric scale; infer other dimensions and layout from
41
+ the video. The anchor's known real dimensions let you work out every other
42
+ object's size and the distances between objects, for example by measuring them
43
+ against the anchor in frames where both are visible, or by scaling a
44
+ reconstruction so that the anchor has its true size. Place the anchor object
45
+ where it is in the video; the room is not aligned to it alone.
46
+
47
+ Visual objective: MAXIMIZE THE VISUAL QUALITY of the reconstructed room, without giving up any
48
+ sim-ready requirement: a scene that looks right but cannot be simulated has failed.
49
+ Use source-aligned renders to improve what a viewer actually sees: recognizable
50
+ furniture silhouettes, correct proportions and placement, detailed surfaces,
51
+ plausible texture scale, window appearance, soft lighting and small distinctive
52
+ objects. Avoid a room made only of plain boxes when the video shows richer
53
+ shapes and materials. Create local textures or meshes when they improve fidelity;
54
+ do not use a single photograph or video plane as a substitute for 3D geometry.
55
+ Check the entire camera path, including corners and reverse views, rather than
56
+ optimizing one attractive render.
57
+
58
+ Required self-critique and refinement:
59
+ 1. Inspect at least 12 frames distributed across the video. Choose at least
60
+ six representative reference views covering the room and important objects.
61
+ Record their timestamps. Estimate corresponding cameras using ONLY the
62
+ public video and scale anchor, and reuse those cameras across iterations.
63
+ 2. Build an early loadable baseline and render those views. Do not spend the
64
+ whole budget on a first draft or a single attractive camera angle.
65
+ 3. Complete at least THREE render-critique-refine-verify rounds. In each round,
66
+ open side-by-side source/render contact sheets with the image-capable Read
67
+ tool. Identify concrete visible discrepancies, implement the highest-impact
68
+ fixes, then re-render the SAME views and check for regressions. Writing a
69
+ critique without changing and re-rendering the scene is not a completed round.
70
+ 4. Round 1 prioritizes room shape, scale, furniture placement, object dimensions
71
+ and silhouettes. Round 2 prioritizes materials, textures, lighting, colors,
72
+ windows and salient details. Round 3 revisits visual mismatches and verifies
73
+ collision support, movable parts and soft items, and runs the self-check.
74
+ 5. Continue additional rounds while time permits. Do not wait until the deadline
75
+ to save a valid scene. Report the true number of completed rounds.
76
+ 6. After each round, save a valid checkpoint and record the source timestamps,
77
+ render paths, specific discrepancies, actual edits, before/after observations,
78
+ and remaining issues in refinement_log.json. Treat self-assessments as
79
+ diagnostics.
80
+ 7. Keep the best validated complete scene, and restore it if a later change
81
+ regresses. The final output must describe that selected result.
82
+
83
+ You may use any tools, including web search, and the software installed here,
84
+ including Blender (`blender`). Download nothing from the internet (no files,
85
+ models, assets, datasets or packages), and do not look up this room, its video
86
+ or the dataset it comes from. Use no annotations, scans or other data about this
87
+ room beyond your inputs, no previous generated outputs and no other session's
88
+ reconstruction. Do not read authentication files. Do not add a robot.
89
+
90
+ The result must be a genuinely interactive 3D scene:
91
+ - Preserve native object identities, count, dimensions, and spatial relationships.
92
+ - Provide collision geometry consistent with visible surfaces and free spaces.
93
+ - Separate movable objects and articulated parts, with appropriate joints and limits.
94
+ - Build every visible soft item with the simulator's native deformable mechanism, as one object
95
+ entry each, with a label that says plainly what it is, for example "curtain".
96
+ - Preserve openings, container interiors, support surfaces, and access needed
97
+ for interactions. Do not replace the room's own objects with generic stand-ins.
98
+ - Assign physically plausible properties and document estimates for unknowns;
99
+ the video is not measured mass, friction or stiffness.
100
+
101
+ Construction rules (the self-check reports them):
102
+ - The scene must start at rest without interpenetration: no contact deeper than
103
+ 5 mm between any two bodies or between a body and fixed geometry.
104
+ - Articulated parts (drawers, doors, lids) must not intersect their own carcass in
105
+ any position of their range: hollow the carcass or exclude the pair from collision.
106
+ - Door and drawer swing paths must be clear of other furniture.
107
+ - The timestep is your choice.
108
+
109
+ Object labels: each objects.json entry has a "label" that says plainly what the object is
110
+ (for example "double bed", "floor lamp"). Use one entry per physical instance.
111
+
112
+ Interaction points: for articulated furniture give a "handle" point on each moving
113
+ part; for movable graspable objects a "grasp" point; for pourable containers a
114
+ "spout" point; for a visible bed cover, four "material_corners" in boundary order.
115
+
116
+ Self-check: run it on your output after each round and before submitting. It loads
117
+ only your output, simulates 3 s from the authored state and reports contacts deeper
118
+ than 5 mm (with the two bodies involved), the time above 5 mm, peak body speed,
119
+ non-finite state and entries without a label. The scene must stay within these
120
+ limits: at most 0.02 s above 5 mm, peak speed at most 5 m/s, finite state. Fix what it reports; do not edit it.
121
+ /isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check
122
+
123
+ Output format (Isaac Sim, in /work/output):
124
+ - scene.usda (or scene.usd): self-contained USD stage, Z-up, metres, explicit default
125
+ prim, locally referenced assets and textures, and an authored PhysX scene.
126
+ - build_scene.py: reproducible builder using the installed Isaac Sim / pxr APIs.
127
+ - objects.json: {"objects": [...]}, one entry per physical object instance:
128
+ "id": a persistent unique string (keep the scale anchor ID),
129
+ "label": what the object is, in plain words,
130
+ "prim_path": the object's root prim (no single mesh containing all furniture),
131
+ "joint_paths": articulated joint prims, if present,
132
+ "interaction_points": interaction Xform prim paths, named handle, grasp or spout
133
+ (child Xforms of the moving part or object),
134
+ "material_corners": four mesh vertex indices for a deformable bed cover.
135
+ - render_preview.py, at least six views rendered with RTX PathTracing, preview.mp4
136
+ (camera tour) and interaction.mp4 (native physics interaction with an original object).
137
+ - README.md and refinement_log.json, as above.
138
+ Runtime: /isaac-sim/python.sh YOUR_SCRIPT.py runs the bundled Python with Isaac Sim,
139
+ USD, NumPy and Pillow; /usr/local/bin/ffmpeg and blender are available. Use headless=True,
140
+ multi_gpu=False, active_gpu=0, physics_gpu=0. sdk_smoke.py is an API and rendering
141
+ example. Native RTX capture can need several app updates before RGB has shape
142
+ (height, width, 4). Only input and output are host-mounted. Do not access
143
+ authentication.
scenes/isaac/47331524/20261114T081855Z/input/scale_anchor.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "object_id": "scale_anchor",
3
+ "description": "The stool outlined in red in the marked image.",
4
+ "reference_image": "anchor_reference.png",
5
+ "marked_image": "anchor_marked.png",
6
+ "region_xyxy_pixels": [
7
+ 158.55090286972978,
8
+ 89.35109928731454,
9
+ 245.7067610769665,
10
+ 185.8587401803795
11
+ ],
12
+ "region_type": "projected_3d_box_envelope_not_segmentation",
13
+ "size_local_xyz_m": [
14
+ 0.33370427970979016,
15
+ 0.3534851293272629,
16
+ 0.4539529003662027
17
+ ],
18
+ "axis_description": [
19
+ "horizontal depth",
20
+ "horizontal width",
21
+ "vertical height"
22
+ ],
23
+ "measurement_type": "full extents of an oriented box around the object",
24
+ "uncertainty_m": null,
25
+ "uncertainty_note": "No uncertainty is supplied for these dimensions."
26
+ }
scenes/isaac/47331524/20261114T081855Z/input/sdk_smoke.py ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Verify Isaac Sim headless RTX rendering and native rigid-body stepping."""
2
+ from isaacsim import SimulationApp
3
+ app = SimulationApp({'headless':True,'renderer':'PathTracing','samples_per_pixel_per_frame':32,
4
+ 'width':512,'height':384,'multi_gpu':False,'active_gpu':0,'physics_gpu':0,
5
+ 'enable_crashreporter':False,'limit_cpu_threads':8})
6
+ import json
7
+ from pathlib import Path
8
+ import numpy as np
9
+ from PIL import Image
10
+ import omni.usd
11
+ import omni.replicator.core as rep
12
+ rep.orchestrator.set_capture_on_play(False)
13
+ from pxr import UsdGeom,UsdLux,UsdShade,UsdPhysics,Gf,Sdf
14
+ stage=omni.usd.get_context().get_stage()
15
+ UsdGeom.SetStageUpAxis(stage,UsdGeom.Tokens.z);UsdGeom.SetStageMetersPerUnit(stage,1.)
16
+ UsdGeom.Xform.Define(stage,'/World');stage.SetDefaultPrim(stage.GetPrimAtPath('/World'))
17
+ def box(path,pos,scale,color,dynamic=False):
18
+ prim=UsdGeom.Cube.Define(stage,path);prim.CreateSizeAttr(1.)
19
+ xf=UsdGeom.Xformable(prim);xf.AddTranslateOp().Set(Gf.Vec3d(*pos));xf.AddScaleOp().Set(Gf.Vec3f(*scale))
20
+ UsdPhysics.CollisionAPI.Apply(prim.GetPrim())
21
+ if dynamic:
22
+ UsdPhysics.RigidBodyAPI.Apply(prim.GetPrim());UsdPhysics.MassAPI.Apply(prim.GetPrim()).CreateMassAttr(1.)
23
+ mat=UsdShade.Material.Define(stage,path+'_Material');shader=UsdShade.Shader.Define(stage,path+'_Material/Shader');shader.CreateIdAttr('UsdPreviewSurface')
24
+ shader.CreateInput('diffuseColor',Sdf.ValueTypeNames.Color3f).Set(Gf.Vec3f(*color));shader.CreateInput('roughness',Sdf.ValueTypeNames.Float).Set(.28)
25
+ mat.CreateSurfaceOutput().ConnectToSource(shader.ConnectableAPI(),'surface');UsdShade.MaterialBindingAPI.Apply(prim.GetPrim()).Bind(mat)
26
+ return prim
27
+ box('/World/Floor',(0,0,-.05),(5,5,.1),(.6,.6,.6));box('/World/Cube',(0,0,1.),(.3,.3,.3),(.08,.3,.8),True)
28
+ light=UsdLux.DomeLight.Define(stage,'/World/Dome');light.CreateIntensityAttr(500.)
29
+ light=UsdLux.RectLight.Define(stage,'/World/Key');light.CreateIntensityAttr(2500.);light.CreateWidthAttr(3.);light.CreateHeightAttr(3.);UsdGeom.Xformable(light).AddTranslateOp().Set(Gf.Vec3d(0,0,3))
30
+ scene=UsdPhysics.Scene.Define(stage,'/World/PhysicsScene');scene.CreateGravityDirectionAttr(Gf.Vec3f(0,0,-1));scene.CreateGravityMagnitudeAttr(9.81)
31
+ camera=rep.create.camera(position=(2.4,-3,2),look_at=(0,0,.3));product=rep.create.render_product(camera,(512,384));rgb=rep.AnnotatorRegistry.get_annotator('rgb');rgb.attach([product])
32
+ from isaacsim.core.api import World
33
+ world=World(physics_dt=1/60,rendering_dt=1/60,stage_units_in_meters=1.)
34
+ world.reset()
35
+ for _ in range(30):world.step(render=True)
36
+ for attempt in range(30):
37
+ rep.orchestrator.step(rt_subframes=8)
38
+ app.update()
39
+ image=np.asarray(rgb.get_data())
40
+ print('CAPTURE_ATTEMPT',attempt,'SHAPE',image.shape,flush=True)
41
+ if image.ndim==3 and image.shape[0]>0:break
42
+ print('RGB_SHAPE',image.shape,flush=True)
43
+ assert image.ndim==3,'RTX annotator did not return a frame'
44
+ position=UsdGeom.Xformable(stage.GetPrimAtPath('/World/Cube')).ComputeLocalToWorldTransform(0).ExtractTranslation()
45
+ print('CUBE_POSITION',list(position),flush=True)
46
+ out=Path('/work/output');out.mkdir(exist_ok=True)
47
+ Image.fromarray(image[:,:,:3]).save(out/'smoke.png');stage.GetRootLayer().Export(str(out/'sdk_example.usda'))
48
+ (out/'smoke.json').write_text(json.dumps({'renderer':'Isaac Sim RTX PathTracing','shape':list(image.shape),'cube_final_position':list(position),'initial_height_m':1.,'success':True}))
49
+ app.close()
scenes/isaac/47331524/20261114T081855Z/input/self_check.py ADDED
@@ -0,0 +1,228 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Passive physics of a USD scene in Isaac Sim, against the scene requirements (no penetration beyond 5 mm for more than
2
+ 0.02 s, no body faster than 5 m/s, finite state).
3
+
4
+ Runs inside the Isaac Sim container. Reads --scene (default /work/scene, read-only) and the optional
5
+ --out/request.json (object ids to drop); writes --out/physics.json (default /work/physics). Without a request it is
6
+ the self-check: no drops, plus a check that every object has a non-empty label.
7
+
8
+ - Speed: a 3 s passive rollout with no rendering and no state reads, timed around world.step only
9
+ (simulated seconds / wall seconds).
10
+ - Passive rollout: 3 s from the authored state. Penetration comes from the PhysX tensor contact view
11
+ (GPU contact reports return nothing under GPU dynamics); speed and poses from the RigidPrim view.
12
+ - Drops: the requested objects with free (unjointed) dynamic bodies, raised 0.05 m
13
+ and released for 2 s, each from a fresh reset.
14
+ The limits are the same as for MuJoCo scenes.
15
+ """
16
+ from isaacsim import SimulationApp
17
+ app=SimulationApp({'headless':True,'multi_gpu':False,'active_gpu':0,'physics_gpu':0,
18
+ 'enable_crashreporter':False,'limit_cpu_threads':8})
19
+ import argparse
20
+ import json
21
+ from pathlib import Path
22
+ import time
23
+ import carb
24
+ import numpy as np
25
+ import omni.usd
26
+ from pxr import UsdPhysics,PhysxSchema
27
+
28
+ LIMIT_M,LIMIT_S,SPEED=.005,.02,5.
29
+ parser=argparse.ArgumentParser();parser.add_argument('--scene',default='/work/scene');parser.add_argument('--out',default='/work/physics')
30
+ args,_=parser.parse_known_args()
31
+ out=Path(args.out);out.mkdir(parents=True,exist_ok=True)
32
+ request=json.loads((out/'request.json').read_text()) if (out/'request.json').is_file() else {'object_ids':[]}
33
+ result={'status':'failed','engine':'Isaac Sim PhysX (GPU dynamics)','no_vlm':True,'provisional':True,
34
+ 'thresholds':{'penetration_m':LIMIT_M,'penetration_duration_s':LIMIT_S,'linear_speed_mps':SPEED},
35
+ 'thresholds_calibrated_across_engines':False}
36
+
37
+
38
+ def contact_filters(stage,paths):
39
+ """GPU contact views filter on rigid bodies or standalone colliders only, so a collider inside a
40
+ rigid body is replaced by that body (unsupported filters are dropped with only a warning)."""
41
+ out=[]
42
+ for path in paths:
43
+ if any('deformable' in str(a).lower() for a in stage.GetPrimAtPath(path).GetAppliedSchemas()):
44
+ continue # deformable meshes cannot be contact filters; their contacts are not measured
45
+ prim=stage.GetPrimAtPath(path);owner=None
46
+ while prim and not prim.IsPseudoRoot():
47
+ if prim.HasAPI(UsdPhysics.RigidBodyAPI):owner=str(prim.GetPath())
48
+ prim=prim.GetParent()
49
+ target=owner or path
50
+ if target not in out:out.append(target)
51
+ return out
52
+
53
+
54
+ def array(value):
55
+ return np.asarray(value.cpu() if hasattr(value,'cpu') else value,dtype=float)
56
+
57
+
58
+ try:
59
+ root=Path(args.scene)
60
+ scene=next(root/n for n in ['scene.usda','scene.usd'] if (root/n).is_file())
61
+ omni.usd.get_context().open_stage(str(scene));stage=omni.usd.get_context().get_stage()
62
+ for _ in range(8):app.update()
63
+
64
+ def dynamic(prim):
65
+ api=UsdPhysics.RigidBodyAPI(prim)
66
+ if api.GetRigidBodyEnabledAttr().Get() not in (None,True) or api.GetKinematicEnabledAttr().Get():return False
67
+ parent=prim.GetParent()
68
+ while parent and not parent.IsPseudoRoot():
69
+ if parent.HasAPI(UsdPhysics.RigidBodyAPI):return False
70
+ parent=parent.GetParent()
71
+ return True
72
+ rigid=[p for p in stage.Traverse() if p.HasAPI(UsdPhysics.RigidBodyAPI)]
73
+ bodies=[p for p in rigid if dynamic(p)]
74
+ paths=[str(p.GetPath()) for p in bodies]
75
+ colliders=[str(p.GetPath()) for p in stage.Traverse() if p.HasAPI(UsdPhysics.CollisionAPI)]
76
+ jointed=set()
77
+ for prim in stage.Traverse():
78
+ if prim.IsA(UsdPhysics.Joint):
79
+ joint=UsdPhysics.Joint(prim)
80
+ for rel in (joint.GetBody0Rel(),joint.GetBody1Rel()):
81
+ jointed.update(str(t) for t in rel.GetTargets())
82
+ for prim in bodies:PhysxSchema.PhysxContactReportAPI.Apply(prim).CreateThresholdAttr().Set(0.)
83
+ scenes=[p for p in stage.Traverse() if p.IsA(UsdPhysics.Scene)]
84
+ authored=PhysxSchema.PhysxSceneAPI(scenes[0]).GetTimeStepsPerSecondAttr().Get() if scenes else None
85
+ rate=float(authored or 60)
86
+ result.update(dynamic_bodies=len(paths),colliders=len(colliders),
87
+ excluded_bodies=[str(p.GetPath()) for p in rigid if not dynamic(p)],
88
+ timestep_s=1/rate,timestep_source='authored PhysxSceneAPI timeStepsPerSecond' if authored else 'default 60 Hz')
89
+ if not paths:raise ValueError('No dynamic rigid bodies to simulate')
90
+ filters=contact_filters(stage,colliders)
91
+ result['contact_filters']=len(filters)
92
+
93
+ carb.settings.get_settings().set_bool('/physics/disableContactProcessing',False)
94
+ from isaacsim.core.api import World
95
+ from isaacsim.core.prims import RigidPrim
96
+ from isaacsim.core.simulation_manager import SimulationManager
97
+ world=World(physics_dt=1/rate,rendering_dt=1/rate,stage_units_in_meters=1.,
98
+ physics_prim_path=str(scenes[0].GetPath()) if scenes else '/World/AddedPhysicsScene',
99
+ set_defaults=False,backend='torch',device='cuda:0')
100
+ context=world.get_physics_context()
101
+ context.enable_gpu_dynamics(True);context.set_broadphase_type('GPU')
102
+ view=RigidPrim(paths,name='scene_bodies');world.scene.add(view)
103
+ steps=lambda seconds:int(np.ceil(seconds*rate))
104
+
105
+ def measure_rtf(dt,repeats=3):
106
+ """Simulated seconds / wall seconds of world.step(render=False) for an unread passive rollout from the authored
107
+ state, median over `repeats` after 10 warm-up steps."""
108
+ world.set_simulation_dt(physics_dt=dt,rendering_dt=dt)
109
+ world.reset()
110
+ for _ in range(10):world.step(render=False)
111
+ count=int(np.ceil(3/dt));samples=[];finite=True
112
+ for _ in range(repeats):
113
+ world.reset();start=time.perf_counter()
114
+ for _ in range(count):world.step(render=False)
115
+ samples.append(count*dt/(time.perf_counter()-start))
116
+ positions,_=view.get_world_poses()
117
+ finite&=bool(np.isfinite(array(positions)).all() and np.isfinite(array(view.get_velocities())).all())
118
+ return {'rtf':float(np.median(samples)),'rtf_samples':samples,'timestep_s':dt,'simulated_seconds':count*dt,
119
+ 'steps':count,'repeats':repeats,'finite':finite,
120
+ 'definition':'simulated seconds / wall seconds of world.step(render=False) only, median of repeats'}
121
+
122
+ result['realtime']={**measure_rtf(1/rate),'scope':'3 s passive rollout from the authored state, authored timestep'}
123
+ print('PHYSICS_STAGE realtime',result['realtime']['rtf'],flush=True)
124
+
125
+ def reset():
126
+ world.reset()
127
+ contacts=SimulationManager.get_physics_sim_view().create_rigid_contact_view(
128
+ paths,filter_patterns=[filters]*len(paths),max_contact_data_count=200000)
129
+ if contacts is None:raise RuntimeError('PhysX rigid contact view could not be created')
130
+ return contacts
131
+
132
+ def rollout(seconds,contacts,motion=None):
133
+ """One rollout; `motion` names the file the logged poses are written to."""
134
+ report={'finite':True,'peak_penetration_m':0.,'penetration_time_s':0.,'peak_linear_speed_mps':0.}
135
+ every=max(1,round(rate/10));frames=[]
136
+ for step in range(steps(seconds)):
137
+ world.step(render=False)
138
+ positions,orientations=view.get_world_poses();velocities=array(view.get_velocities())
139
+ if step%every==0:frames.append(((step+1)/rate,array(positions),array(orientations)))
140
+ report['finite']&=bool(np.isfinite(array(positions)).all() and np.isfinite(velocities).all())
141
+ _,_,_,separation,counts,starts=contacts.get_contact_data(dt=1/rate)
142
+ total=int(array(counts).sum())
143
+ if 'initial_pairs' not in report:
144
+ # Pairs deeper than the limit after the first step, named for the self-check.
145
+ pairs=[];sep=array(separation);cnt=array(counts).astype(int);beg=array(starts).astype(int)
146
+ for i,j in zip(*np.nonzero(cnt)):
147
+ depth=-float(sep[beg[i,j]:beg[i,j]+cnt[i,j]].min())
148
+ if depth>LIMIT_M:pairs.append({'depth_m':depth,'bodies':[paths[i],filters[j]]})
149
+ report['initial_pairs']=sorted(pairs,key=lambda p:-p['depth_m'])[:20]
150
+ deepest=max(0.,-float(array(separation)[:total].min())) if total else 0.
151
+ report['peak_penetration_m']=max(report['peak_penetration_m'],deepest)
152
+ if deepest>LIMIT_M:report['penetration_time_s']+=1/rate
153
+ report['peak_linear_speed_mps']=max(report['peak_linear_speed_mps'],float(np.linalg.norm(velocities[:,:3],axis=1).max()))
154
+ if not report['finite']:break
155
+ report['valid']=bool(report['finite'] and report['penetration_time_s']<=LIMIT_S and report['peak_linear_speed_mps']<=SPEED)
156
+ report['thresholds']=result['thresholds']
157
+ if motion and request.get('record'):
158
+ (out/'motion').mkdir(exist_ok=True)
159
+ np.savez_compressed(out/'motion'/motion,paths=np.array(paths),t=np.array([f[0] for f in frames]),
160
+ positions=np.stack([f[1] for f in frames]),orientations=np.stack([f[2] for f in frames]))
161
+ report['motion']=f'motion/{motion}.npz'
162
+ return report
163
+
164
+ contacts=reset()
165
+ start_positions=array(view.get_world_poses()[0])
166
+ masses=array(view.get_masses())
167
+ direction,magnitude=context.get_gravity()
168
+ gravity=np.asarray(direction,dtype=float)*float(magnitude)
169
+ passive=rollout(3,contacts,'passive')
170
+ print('PHYSICS_STAGE passive',passive,flush=True)
171
+ end_positions=array(view.get_world_poses()[0])
172
+ moved=np.linalg.norm(end_positions-start_positions,axis=1)
173
+
174
+ manifest=json.loads((root/'objects.json').read_text())
175
+ manifest=manifest['objects'] if isinstance(manifest,dict) else manifest
176
+ result['manifest_errors']=[f"{o.get('id')}: label is missing or empty" for o in manifest if not str(o.get('label') or '').strip()]
177
+ result['manifest_errors']+=[f"{o.get('id')}: prim_path does not resolve" for o in manifest if not stage.GetPrimAtPath(o.get('prim_path','') or '/')]
178
+ def owned(obj):
179
+ prefix=str(obj.get('prim_path','')).rstrip('/')
180
+ return [i for i,p in enumerate(paths) if prefix and (p==prefix or p.startswith(prefix+'/'))]
181
+ drift=[{'id':o['id'],'dynamic_bodies':len(owned(o)),'center_drift_m':float(moved[owned(o)].max()),'mass_kg':float(masses[owned(o)].sum())}
182
+ for o in manifest if owned(o)]
183
+
184
+ requested=set(request['object_ids'])
185
+ eligible=[o for o in manifest if o['id'] in requested]
186
+ drops=[]
187
+ for obj in eligible:
188
+ free=[i for i in owned(obj) if paths[i] not in jointed]
189
+ if not free:continue
190
+ contacts=reset()
191
+ positions,orientations=view.get_world_poses()
192
+ raised=positions.clone() if hasattr(positions,'clone') else positions.copy()
193
+ raised[free,2]+=.05
194
+ view.set_world_poses(positions=raised,orientations=orientations)
195
+ applied=float((array(view.get_world_poses()[0])[free,2]-array(positions)[free,2]).min())
196
+ drops.append({'id':obj['id'],'input':'raise free bodies 0.05 m and release','bodies':[paths[i] for i in free],
197
+ 'lift_applied_m':applied,**rollout(2,contacts,f'drop_{len(drops)}')})
198
+ print('PHYSICS_STAGE drop',obj['id'],applied,drops[-1]['peak_linear_speed_mps'],flush=True)
199
+
200
+ # The common 2 ms step last, so the rollouts above run at the authored step.
201
+ result['realtime_common_step']={**measure_rtf(.002),'scope':'3 s passive rollout from the authored state at the common 2 ms step'}
202
+ print('PHYSICS_STAGE realtime_common_step',result['realtime_common_step']['rtf'],flush=True)
203
+ gravity_ok=bool(np.allclose(gravity,[0,0,-9.81],atol=.1))
204
+ mass_ok=bool(np.isfinite(masses).all() and (masses>=0).all())
205
+ result.update(status='completed',passive=passive,gravity_m_s2=gravity.tolist(),gravity_valid=gravity_ok,
206
+ mass_valid=mass_ok,center_drift=drift,drop_tests=drops,eligible_native_objects=len(eligible),
207
+ largest_body_displacements=[{'prim':paths[i],'displacement_m':float(moved[i])} for i in np.argsort(-moved)[:5]],
208
+ valid=bool(passive['valid'] and gravity_ok and mass_ok),
209
+ drop_scope='the requested objects with free (unjointed) dynamic bodies, in manifest order',
210
+ contact_scope=('contacts between each dynamic body and every collider (static, kinematic and dynamic; colliders '
211
+ 'inside a rigid body are filtered by that body, as GPU contact views require); '
212
+ 'static-static pairs are not simulated, as in MuJoCo world-body geoms'),
213
+ limitations=['the limits are the MuJoCo ones, not calibrated for PhysX',
214
+ 'kinematic, disabled and nested rigid bodies are not traced (see excluded_bodies)',
215
+ 'drift is per dynamic body under the object prim, not the object bounding box',
216
+ 'contacts with deformable bodies are not measured (GPU contact views cannot filter them)'])
217
+ except Exception as error:
218
+ result['error']=f'{type(error).__name__}: {error}'
219
+ finally:
220
+ (out/'physics.json').write_text(json.dumps(result,indent=2)+'\n')
221
+ if 'passive' in result:
222
+ passive=result['passive']
223
+ print('PASSIVE CHECK','PASSED' if passive['valid'] and not result.get('manifest_errors') else 'FAILED',
224
+ f"- {len(passive.get('initial_pairs',[]))} contact pairs > 5 mm, {passive['penetration_time_s']:.3f} s above 5 mm,"
225
+ f" peak speed {passive['peak_linear_speed_mps']:.2f} m/s, {len(result.get('manifest_errors',[]))} manifest errors;"
226
+ f" details in {out/'physics.json'}",flush=True)
227
+ else:print('CHECK FAILED:',result.get('error'),flush=True)
228
+ app.close()
scenes/isaac/47331524/20261114T081855Z/input/video.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9ae6725724ceeffa029290fbb935e9b3e7898120b6b9b3b5af5383584c718dd7
3
+ size 5598485
scenes/isaac/47331524/20261114T081855Z/run.json ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "track": "scenes",
3
+ "engine": "isaac",
4
+ "item": "47331524",
5
+ "run_id": "20261114T081855Z",
6
+ "campaign": "jiarui-2-codebuddy",
7
+ "cli": "codebuddy",
8
+ "model": "Hunyuan 4",
9
+ "model_id": "hy4-preview",
10
+ "model_family": "hy4-preview",
11
+ "model_reasoning_effort": "default",
12
+ "budget_seconds": 900,
13
+ "told_budget_seconds": 900,
14
+ "stage": 1,
15
+ "continued_from": null,
16
+ "contract": {
17
+ "name": "scenes-isaac",
18
+ "sha256": "f1ab7ea88842502d37176b41d9f3bec64a5d9fac292ca3f16caf08da3b5db96d",
19
+ "track": "scenes",
20
+ "engine": "isaac",
21
+ "engine_version": "6.2.0",
22
+ "image": "simready-isaac-generation:latest",
23
+ "blender_version": "5.2.3",
24
+ "inputs": [
25
+ "video.mp4",
26
+ "scale_anchor.json",
27
+ "anchor_reference.png",
28
+ "anchor_marked.png",
29
+ "prompt.txt",
30
+ "generation_spec.json",
31
+ "self_check.py",
32
+ "sdk_smoke.py"
33
+ ],
34
+ "withheld": [
35
+ "groundtruth/ (objects, cameras, task catalogs, object dynamics, interaction test) and its source tag",
36
+ "provenance/ (source annotation, anchor selection)",
37
+ "the frozen protocol of the track and engine (data/benchmark/frozen/<track>/<engine>.json)",
38
+ "other runs"
39
+ ],
40
+ "outputs": [
41
+ "scene.usda",
42
+ "build_scene.py",
43
+ "objects.json",
44
+ "render_preview.py",
45
+ "preview.mp4",
46
+ "interaction.mp4",
47
+ "README.md",
48
+ "refinement_log.json"
49
+ ],
50
+ "self_check": "/isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check",
51
+ "budget": {
52
+ "told": "the first of the session's cumulative budgets, in wall-clock seconds from container start",
53
+ "resume": "each later budget resumes the same conversation for the rest of its time; each budget is its own run",
54
+ "enforcement": "the runner kills the container at the deadline; the output as it stands is the result"
55
+ },
56
+ "refinement": {
57
+ "minimum_completed_rounds": 3,
58
+ "minimum_reference_views": 6,
59
+ "minimum_inspected_video_frames": 12,
60
+ "final_validation_seconds": 90
61
+ },
62
+ "isolation": {
63
+ "container": "Docker: read-only input (and tools), writable output and scratch, temporary credentials; no repository, benchmark ground truth or evaluator files",
64
+ "network": "an internal network whose only exit is an HTTPS CONNECT proxy (runtime/egress_proxy.py)",
65
+ "egress_hosts": [
66
+ "api.anthropic.com",
67
+ "platform.claude.com",
68
+ "chatgpt.com",
69
+ "auth.openai.com",
70
+ "opencode.ai",
71
+ "mcp.exa.ai",
72
+ "search.parallel.ai"
73
+ ]
74
+ },
75
+ "protocol": {
76
+ "file": "data/benchmark/frozen/scenes/isaac.json",
77
+ "sha256": "df0fe5bb1c796474ad13cdba81b660a2946aa24297cb289f9581cfde7c035bb7"
78
+ }
79
+ },
80
+ "engine_version": "6.2.0",
81
+ "image": "simready-isaac-generation:latest",
82
+ "image_id": "sha256:8d9f9c934457aa3c8e91e8099e35db27b50db287d0eb76a647052c310e754d87",
83
+ "required_refinement_rounds": 3,
84
+ "gpu_uuid": "GPU-91246d27-fecc-875a-e18a-afb185bc6a18",
85
+ "egress_mode": "internal_network",
86
+ "input_snapshot": "input",
87
+ "tools": null,
88
+ "evaluator_protocol": "evaluator_protocol.json",
89
+ "output": "output",
90
+ "log": "generation.jsonl",
91
+ "status": "budget_expired",
92
+ "agent_cli": {
93
+ "cli": "codebuddy",
94
+ "name": "CodeBuddy Code",
95
+ "binary": "a pinned codebuddy-code package and its node, mounted read-only (SIMREADY_CODEBUDDY_PREFIX, default the host's codebuddy install)",
96
+ "login": "the CodeBuddy login file, given to the CLI as a snapshot without its refresh token (the session cannot rotate it)",
97
+ "billing": "CodeBuddy account (the CLI reports no price)",
98
+ "tools": "every tool but WebFetch and the server-side media generators ImageGen, ImageEdit, VideoGen (--disallowedTools); WebSearch runs on CodeBuddy's side",
99
+ "hosts": [
100
+ "[redacted-host]",
101
+ "wb.[redacted]buddy.com"
102
+ ],
103
+ "version": "2.182.0",
104
+ "model": "hy4-preview",
105
+ "model_family": "hy4-preview"
106
+ },
107
+ "started_at": "2026-11-14T08:18:57.381831+00:00",
108
+ "model_answered": true,
109
+ "exit_code": 137,
110
+ "credential_refresh_persistence": "snapshot_without_refresh_token",
111
+ "finished_at": "2026-11-14T08:33:58.822659+00:00",
112
+ "cli_session_id": "01a1d923-6d50-74db-a82d-57b3ce1f5afe",
113
+ "api_retries": 0,
114
+ "rate_limit_events": 0,
115
+ "stage_seconds": 901.451,
116
+ "elapsed_seconds": 901.451,
117
+ "cost": {
118
+ "status": "partial",
119
+ "source": "the CLI's running usage totals (usage_update events)",
120
+ "usd": null,
121
+ "usd_basis": "not reported (CodeBuddy account billing)",
122
+ "credits": 0,
123
+ "credit_status": "known",
124
+ "turns": null,
125
+ "api_seconds": null,
126
+ "tokens": {
127
+ "input": 56429,
128
+ "cache_read": 326848,
129
+ "cache_write": 0,
130
+ "output": 2445,
131
+ "reasoning": null,
132
+ "total": 385722
133
+ },
134
+ "wall_seconds": 901.451,
135
+ "budget_seconds": 900,
136
+ "stopped_at_budget": true
137
+ },
138
+ "scene_present": false
139
+ }
scenes/isaac/47331524/20261114T081855Z/submitted_prompt.txt ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ About the simulator. NVIDIA Isaac Sim is a robotics simulator built on Omniverse. You describe a scene
2
+ as an OpenUSD stage: prims carrying physics schemas for rigid bodies, colliders, joints and articulations, and
3
+ deformable bodies and surfaces. NVIDIA PhysX simulates it, here with GPU dynamics, and RTX path tracing renders it.
4
+ Isaac Sim is widely used for robot learning and sim-to-real work.
5
+
6
+ Reconstruct the room shown in video.mp4 as an interactive Isaac Sim scene.
7
+
8
+ THE RESULT MUST BE SIM-READY, READY FOR INTERACTIVE ROBOT LEARNING. It is an environment a robot learns
9
+ in, not a picture of a room. A training pipeline must be able to load it headless, reset it, step it for
10
+ a long time and let robots act on it, with no manual repair. That means:
11
+ - It loads with no network access or external files, and resets to the same state at rest.
12
+ - Everything a robot could touch has collision geometry matching what is visible. An object with no
13
+ collision does not exist for a robot, however good it looks.
14
+ - Rigid things are rigid bodies with plausible mass; free objects can be pushed, and doors, drawers and lids
15
+ move on joints with limits.
16
+ - Soft things are native deformables: cloth, bedding, curtains, rugs, throws, cushions and pillows.
17
+ A rigid body that merely looks soft is wrong.
18
+ - It stays stable under sustained interaction, across many steps without a reset, and steps fast
19
+ enough to train on.
20
+
21
+ Budget: 15 minutes (900 seconds) of wall-clock time per scene, starting when
22
+ you receive access to the input files. All inspection, coding, asset creation,
23
+ tool calls, rendering, debugging, and validation must fit within this budget.
24
+ Stop work at the deadline and submit the latest valid, loadable result.
25
+ Do not stop after producing the first runnable scene. Use the available time
26
+ for repeated self-critique and refinement; reserve the final 90 seconds for
27
+ validation and submission.
28
+
29
+ Inputs (in /work/input):
30
+ - video.mp4: the complete source video.
31
+ - scale_anchor.json: dimensions and identity of one object in the room, the scale anchor.
32
+ - anchor_reference.png and anchor_marked.png: images identifying that object.
33
+ The marked rectangle is a projected box envelope, not a segmentation mask.
34
+ - generation_spec.json: this contract in machine-readable form.
35
+ - self_check.py: the public self-check described below.
36
+
37
+ Maximize visual realism and fidelity to the source video. Reproduce the room's
38
+ geometry, proportions, native objects, materials, textures, colors, lighting,
39
+ and visible details as faithfully as possible within the time budget. Use the
40
+ scale anchor to establish metric scale; infer other dimensions and layout from
41
+ the video. The anchor's known real dimensions let you work out every other
42
+ object's size and the distances between objects, for example by measuring them
43
+ against the anchor in frames where both are visible, or by scaling a
44
+ reconstruction so that the anchor has its true size. Place the anchor object
45
+ where it is in the video; the room is not aligned to it alone.
46
+
47
+ Visual objective: MAXIMIZE THE VISUAL QUALITY of the reconstructed room, without giving up any
48
+ sim-ready requirement: a scene that looks right but cannot be simulated has failed.
49
+ Use source-aligned renders to improve what a viewer actually sees: recognizable
50
+ furniture silhouettes, correct proportions and placement, detailed surfaces,
51
+ plausible texture scale, window appearance, soft lighting and small distinctive
52
+ objects. Avoid a room made only of plain boxes when the video shows richer
53
+ shapes and materials. Create local textures or meshes when they improve fidelity;
54
+ do not use a single photograph or video plane as a substitute for 3D geometry.
55
+ Check the entire camera path, including corners and reverse views, rather than
56
+ optimizing one attractive render.
57
+
58
+ Required self-critique and refinement:
59
+ 1. Inspect at least 12 frames distributed across the video. Choose at least
60
+ six representative reference views covering the room and important objects.
61
+ Record their timestamps. Estimate corresponding cameras using ONLY the
62
+ public video and scale anchor, and reuse those cameras across iterations.
63
+ 2. Build an early loadable baseline and render those views. Do not spend the
64
+ whole budget on a first draft or a single attractive camera angle.
65
+ 3. Complete at least THREE render-critique-refine-verify rounds. In each round,
66
+ open side-by-side source/render contact sheets with the image-capable Read
67
+ tool. Identify concrete visible discrepancies, implement the highest-impact
68
+ fixes, then re-render the SAME views and check for regressions. Writing a
69
+ critique without changing and re-rendering the scene is not a completed round.
70
+ 4. Round 1 prioritizes room shape, scale, furniture placement, object dimensions
71
+ and silhouettes. Round 2 prioritizes materials, textures, lighting, colors,
72
+ windows and salient details. Round 3 revisits visual mismatches and verifies
73
+ collision support, movable parts and soft items, and runs the self-check.
74
+ 5. Continue additional rounds while time permits. Do not wait until the deadline
75
+ to save a valid scene. Report the true number of completed rounds.
76
+ 6. After each round, save a valid checkpoint and record the source timestamps,
77
+ render paths, specific discrepancies, actual edits, before/after observations,
78
+ and remaining issues in refinement_log.json. Treat self-assessments as
79
+ diagnostics.
80
+ 7. Keep the best validated complete scene, and restore it if a later change
81
+ regresses. The final output must describe that selected result.
82
+
83
+ You may use any tools, including web search, and the software installed here,
84
+ including Blender (`blender`). Download nothing from the internet (no files,
85
+ models, assets, datasets or packages), and do not look up this room, its video
86
+ or the dataset it comes from. Use no annotations, scans or other data about this
87
+ room beyond your inputs, no previous generated outputs and no other session's
88
+ reconstruction. Do not read authentication files. Do not add a robot.
89
+
90
+ The result must be a genuinely interactive 3D scene:
91
+ - Preserve native object identities, count, dimensions, and spatial relationships.
92
+ - Provide collision geometry consistent with visible surfaces and free spaces.
93
+ - Separate movable objects and articulated parts, with appropriate joints and limits.
94
+ - Build every visible soft item with the simulator's native deformable mechanism, as one object
95
+ entry each, with a label that says plainly what it is, for example "curtain".
96
+ - Preserve openings, container interiors, support surfaces, and access needed
97
+ for interactions. Do not replace the room's own objects with generic stand-ins.
98
+ - Assign physically plausible properties and document estimates for unknowns;
99
+ the video is not measured mass, friction or stiffness.
100
+
101
+ Construction rules (the self-check reports them):
102
+ - The scene must start at rest without interpenetration: no contact deeper than
103
+ 5 mm between any two bodies or between a body and fixed geometry.
104
+ - Articulated parts (drawers, doors, lids) must not intersect their own carcass in
105
+ any position of their range: hollow the carcass or exclude the pair from collision.
106
+ - Door and drawer swing paths must be clear of other furniture.
107
+ - The timestep is your choice.
108
+
109
+ Object labels: each objects.json entry has a "label" that says plainly what the object is
110
+ (for example "double bed", "floor lamp"). Use one entry per physical instance.
111
+
112
+ Interaction points: for articulated furniture give a "handle" point on each moving
113
+ part; for movable graspable objects a "grasp" point; for pourable containers a
114
+ "spout" point; for a visible bed cover, four "material_corners" in boundary order.
115
+
116
+ Self-check: run it on your output after each round and before submitting. It loads
117
+ only your output, simulates 3 s from the authored state and reports contacts deeper
118
+ than 5 mm (with the two bodies involved), the time above 5 mm, peak body speed,
119
+ non-finite state and entries without a label. The scene must stay within these
120
+ limits: at most 0.02 s above 5 mm, peak speed at most 5 m/s, finite state. Fix what it reports; do not edit it.
121
+ /isaac-sim/python.sh /work/input/self_check.py --scene /work/output --out /tmp/self_check
122
+
123
+ Output format (Isaac Sim, in /work/output):
124
+ - scene.usda (or scene.usd): self-contained USD stage, Z-up, metres, explicit default
125
+ prim, locally referenced assets and textures, and an authored PhysX scene.
126
+ - build_scene.py: reproducible builder using the installed Isaac Sim / pxr APIs.
127
+ - objects.json: {"objects": [...]}, one entry per physical object instance:
128
+ "id": a persistent unique string (keep the scale anchor ID),
129
+ "label": what the object is, in plain words,
130
+ "prim_path": the object's root prim (no single mesh containing all furniture),
131
+ "joint_paths": articulated joint prims, if present,
132
+ "interaction_points": interaction Xform prim paths, named handle, grasp or spout
133
+ (child Xforms of the moving part or object),
134
+ "material_corners": four mesh vertex indices for a deformable bed cover.
135
+ - render_preview.py, at least six views rendered with RTX PathTracing, preview.mp4
136
+ (camera tour) and interaction.mp4 (native physics interaction with an original object).
137
+ - README.md and refinement_log.json, as above.
138
+ Runtime: /isaac-sim/python.sh YOUR_SCRIPT.py runs the bundled Python with Isaac Sim,
139
+ USD, NumPy and Pillow; /usr/local/bin/ffmpeg and blender are available. Use headless=True,
140
+ multi_gpu=False, active_gpu=0, physics_gpu=0. sdk_smoke.py is an API and rendering
141
+ example. Native RTX capture can need several app updates before RGB has shape
142
+ (height, width, 4). Only input and output are host-mounted. Do not access
143
+ authentication.
scenes/isaac/47334756/20261114T061724Z/evaluator_protocol.json ADDED
The diff for this file is too large to render. See raw diff
 
scenes/isaac/47334756/20261114T061724Z/generation.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
scenes/isaac/47334756/20261114T061724Z/input/anchor_marked.png ADDED

Git LFS Details

  • SHA256: b809fae0d8f0cf69f0d6bfd952bec13059d0472e2dcdd553e446308e28f6e34c
  • Pointer size: 130 Bytes
  • Size of remote file: 61.4 kB