Stack Two Cups
50 Franka demonstrations in LeRobot v3.0 format; 19,838 frames at 30 Hz, with three native 672×376 RIGHT-eye RGB videos (wrist, exterior, exterior2). Exterior2 is stored rotated 180° as during collection.
Instructions
- stack two cups
Episode instructions are embedded in meta/tasks.parquet, every data row's
task_index, meta/episodes/, and meta/episode_annotations.json.
Recorded values
action: seven absolute joint command positions in radians, followed by
OPEN=1 / CLOSE=0. observation.state: seven measured joints, gripper width/0.08,
previous commanded gripper latch, and configured-TCP xyz/rotation vector.
Raw 15D state is retained; a training adapter may explicitly select joints and
measured gripper width. Flange state, request/accepted Cartesian diagnostics,
command/segment IDs and native monotonic timestamps are also retained.
Videos and confirmed command traces are byte-identical copies. Actions, states,
frame timestamps and diagnostics are unchanged. Only dataset/episode/task IDs
and saved language labels are remapped across recording sessions. Source
metadata and per-episode execution profiles are preserved in meta/source_sessions/.
Select the episode's contract through meta/action_contract.json's
episode_contracts mapping; profiles can vary between sessions.
All 50 confirmed episodes and their complete recorded frames are included,
including short clutch segments. No training crop, resize, normalization,
trajectory filtering or train/validation split has been applied. Fit any
training normalization on the chosen training split. Failed/unconfirmed traces
are excluded. SHA256SUMS covers this exported copy.
Per-episode reviewed values preserve supplied review status; an automated
data audit does not mark an episode as user reviewed.
- Downloads last month
- 15