erensckin/hand2throw-teleop-demos

Hand2Throw: teleop demos

205 demonstrations of "put the ketchup in the basket" in LIBERO (Franka Panda, MuJoCo), recorded with live hand-tracking teleoperation (webcam + phone, MediaPipe). Code and full write-up: https://github.com/erensckin/hand2throw

The basket stands 0.70, 0.80, 0.90 or 1.00 m in front of the robot. At 0.70 m the ketchup is placed; beyond reach it is thrown. The operator controlled reach, grasp and lift by hand and pressed a key to trigger a scripted throw, whose strength was computed from the true basket distance.

Contents

• Format: LeRobotDataset v3.0 (loadable with LeRobot). 205 episodes, 33,741 frames, 20 fps.
• observation.images.image (agentview), image2 (wrist), image3 (side): 256×256 video, sim renders only.
• observation.state (8-D), action (7-D: OSC_POSE end-effector deltas + gripper), task text.
• extra/episodes.jsonl: per episode the basket distance, strategy (place/throw), throw parameters and object positions.
• extra/exclude.txt: episodes that met the strategy rule but were excluded by hand (see below).

200 of the 205 episodes were used for training (50 per distance). Episodes are selected by scripts/select_episodes.py in the GitHub repo, which keeps successful episodes whose strategy matches the distance (place below 0.775 m, throw beyond), minus extra/exclude.txt. Not used:

• 4, 8, 12: accidental throws at 0.70 m ("strategy": "throw" in extra/episodes.jsonl), dropped by the strategy rule;
• 32, 36: placed successfully but knocked the basket over, listed in extra/exclude.txt.

They are kept so that episode indices stay stable. Each model's train_config.json in hand2throw-smolvla lists its exact episodes.

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("erensckin/hand2throw-teleop-demos")