Datasetpublic

ivogeorg/aic-cable-insertion-gazebo-v3

AIC Dexterous Cable Insertion — Gazebo (LeRobot v3) Simulated demonstration dataset for the dexterous cable-insertion task from the AI for Industry Challenge (AIC) 2026, Qualification phase. Recorded in the Gazebo AIC simulator from an autonomous demonstrator policy and packaged in the LeRobot v3 format. Dataset summary Format LeRobot v3.0 Robot type aic_controller Episodes 274 Frames 289,753 FPS 20 Tasks 1 — "Insert the cable connector into… See the full description on the dataset page: https://huggingface.co/datasets/ivogeorg/aic-cable-insertion-gazebo-v3.

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes46downloads

Dataset Card

AIC Dexterous Cable Insertion — Gazebo (LeRobot v3)

Simulated demonstration dataset for the dexterous cable-insertion task from the AI for Industry Challenge (AIC) 2026, Qualification phase. Recorded in the Gazebo AIC simulator from an autonomous demonstrator policy and packaged in the LeRobot v3 format.

Dataset summary


Format	LeRobot `v3.0`
Robot type	`aic_controller`
Episodes	274
Frames	289,753
FPS	20
Tasks	1 — "Insert the cable connector into the correct target port"
Simulator	Gazebo (AIC `aic_eval` toolkit)

Features

Key	dtype	shape	Description
`observation.images.left_camera`	image	256×288×3	Wrist-mounted left camera
`observation.images.center_camera`	image	256×288×3	Wrist-mounted center camera
`observation.images.right_camera`	image	256×288×3	Wrist-mounted right camera
`observation.state`	float32	(6,)	Arm joint positions (6 DoF)
`observation.environment_state`	float32	(6,)	Wrist force/torque wrench `[fx, fy, fz, tx, ty, tz]`
`action`	float32	(7,)	TCP pose target `[x, y, z, qx, qy, qz, qw]`

Plus the standard LeRobot bookkeeping columns (timestamp, frame_index, episode_index, index, task_index). Per-feature normalization statistics (min / max / mean / std / q01…q99) are in meta/stats.json.

How it was generated

1.An autonomous demonstrator policy drives the arm in the Gazebo AIC simulation (a ground-truth-guided controller — not human teleoperation).
2.`aic_factory_recorder.py` records each trial directly to LeRobot v3 parquet — 3 wrist cameras + joint state + wrist wrench + pose action at 20 Hz.
3.`build_streaming_dataset.py` replays every accepted trial through LeRobot's own LeRobotDataset writer to produce this canonical dataset.

Only trials passing the AIC reject filter (valid Tier-1, successful insertion event) were retained.

Loading

python

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("<user>/<repo-name>")
print(ds.meta.info["total_episodes"], ds.meta.info["total_frames"])

Intended use & limitations

—Built for imitation learning (e.g. ACT) on the AIC cable-insertion task. It is task- and simulator-specific (Gazebo, the AIC task board) — not a general manipulation dataset.
—All demonstrations are synthetic / autonomously generated, not human teleoperation.

Provenance

Generated during the AIC 2026 Qualification phase. The generation pipeline (factory scripts, aic_factory_recorder.py, build_streaming_dataset.py) is maintained in the team's challenge repository.