rkstgr/aic-lewm-gazebo

AIC Cable-Insertion Exploration — Gazebo (LeRobot format)

Exploration dataset from the Intrinsic AI for Industry Challenge (AIC) cable-insertion task, collected in the AIC Gazebo simulation. 500 episodes / 25,000 frames at 10 Hz, intended as pre-training data for a LeWM (Learned World Model).

The robot is a UR5e holding a cable, in a scene with a task board, two SC ports, and a NIC card mount. The exploration policies are deliberately not trying to insert the cable — they are sampling state-action transitions across the workspace, including aggressive contact with the board, so a world model trained on this data sees both free-space and contact dynamics.

Schema

Key	Type	Shape	Notes
observation.images.left	image (uint8)	224 × 224 × 3	Left wrist camera
observation.images.center	image (uint8)	224 × 224 × 3	Center wrist camera
observation.images.right	image (uint8)	224 × 224 × 3	Right wrist camera
observation.state	float32	(32,)	See breakdown below
action	float32	(6,)	Cartesian twist: [lin_x, lin_y, lin_z, ang_x, ang_y, ang_z], capped at ±0.05 m/s linear / ±0.30 rad/s angular
exploration_mode	string	—	One of random_twist, scripted_scan, aggressive_descent

State vector layout (32-D, all in the AIC controller's gripper/tcp frame):

Index	Field	Source topic
0–2	TCP position (x, y, z)	/aic_controller/controller_state
3–6	TCP orientation (qx, qy, qz, qw)	same
7–9	TCP linear velocity	same
10–12	TCP angular velocity	same
13–15	TCP tracking error (linear)	same
16–18	TCP tracking error (angular)	same
19–25	Joint positions (6 UR5e arm + 1 gripper slot)	/joint_states
26–28	F/T sensor force (fx, fy, fz)	/fts_broadcaster/wrench
29–31	F/T sensor torque (tx, ty, tz)	same

Exploration modes

One mode per episode (50 frames). Empirical distribution across the 500 episodes:

• random_twist — 45 % — smoothed random twist (EMA, re-sampled every 8 frames)
• scripted_scan — 30 % — Lissajous-style scan with random per-episode phase
• aggressive_descent — 25 % — push downward toward the board to capture contact / collision dynamics

How the dataset was collected

• AIC Gazebo sim launched with start_aic_engine:=false ground_truth:=true gazebo_gui:=false. No aic_engine, no aic_model node — just the simulation half plus a custom Python collector that publishes to /aic_controller/pose_commands and subscribes to the topics above.
• Soft reset between episodes (P-control toward home pose (0.45, 0.0, 0.45) for 30 frames) keeps the arm in the useful workspace.
• Saved with LeRobotDataset.create(..., use_videos=False) so each frame's three images are stored as PNGs inline in the parquet rows (no video encoding).
• Wallclock for the full run: 86 min on an A4000.

Caveats

• Free-space drift dominates: ~45 % of episodes are purely random twists in free space. If a downstream world model overfits to "drift in empty space," consider reweighting toward aggressive_descent frames.
• All quantities are in the gripper TCP frame, not the world / base frame.
• No CheatCode-style "successful insertion" frames — the dataset has no examples of the cable seated in a port.
• Frame 0 of every episode is near the home pose (low initial-state diversity).

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("rkstgr/aic-lewm-gazebo")
print(ds.num_episodes, ds.num_frames, ds.fps)
print(ds[0].keys())

License

Apache-2.0 (matching the AIC repo / Intrinsic Innovation LLC's example code).