MacroSwan/cablepilot-motion

cablepilot motion

The data that the cablepilot world model (weights/stringman_g1.pt, 28k parameters) learned from. It contains 12,000 task-free trajectories, about 100 hours of simulated motion of a Stringman cable robot with its stock gripper. Each trajectory is 30 s long, recorded at 10 Hz, in its own random room. There is no target, no cost and no task label: only how the machine moves when it is commanded. Tasks come later, as cards (MacroSwan/pilot-tasks).

Files

data/shard_XXXX.npz holds 120 shards of 100 trajectories each, and data/manifest.json records the recipe, the seed and the field names.

key	shape	what
states	(N, 301, 14)	gantry position p and velocity v, pole swing s and its rate, wrist angle and rate, finger closure and rate. These are as estimated (with the estimator's noise), the way the robot would see them
actions	(N, 300, 5)	the command sent: gantry velocity (3, what move_direction_speed takes), wrist rate, closure target
tensions	(N, 301, 4)	line tensions (N). An observation only, not a model input
payload	(N, 301)	kg held at the grasp point. The model has to infer it
eta	(N, P)	this trajectory's randomised physics (names in the manifest)
mounts	(N, 4, 3)	the room's four line anchors, in the firmware's line order. This is how the room reaches the model
runs	(N, 4)	the fixed run length of each indirect line

How it was made

scripts/cdpr_datagen.py in the model repository (its docstring gives every range):

python scripts/cdpr_datagen.py --scenario configs/scenarios/clear_playroom.yaml --n 12000 --seed 0 --workers 4 --out data

The robot is driven the way the real host drives it. Line speeds come from finite-difference IK on the estimated position, through mounts that are miscalibrated on purpose. The spool loop runs at 50 Hz with the tension floor. Motors are acceleration-limited, and commands arrive with latency. The motion is a random chain of the robot's real regimes: go to a point, dwell, stop hard, pump the swing, descend to grasp height, and MPPI-like command noise. Payloads from 20 to 500 g are picked up and dropped. Each trajectory randomises mass, line stiffness and damping, spool acceleration, latency, calibration error and swing damping. Rooms are 3 to 7 m jittered rectangles with mounts 2.0 to 3.2 m high.

This copy was regenerated from that script, seed 0, at the release commit (the original files were not kept). On its last 500 trajectories, which are held out, the released weights score 20.4 mm median grasp-point error at 2 s, against 21.5 mm for the MuJoCo oracle. The original test gave 19.6 mm against 21.0 mm. Both moved by the same half millimetre, which fits the same distribution drawn with a different random sequence.

Use it

Retrain or fine-tune with scripts/train_cdpr_wm.py --data <this folder>. A model that beats the released one on test 1 and on a task is welcome as a PR on MacroSwan/cablepilot (see its AGENTS.md, Step 7). Real Stringman recordings, such as naavox/stringman_cleans, are the natural test set. Nobody has measured the model on real data yet.

Apache-2.0. Simulated, so it has no people and no homes in it.