soarm_amazing_hand_pick

SO-ARM101 + AmazingHand Grasping Dataset

Teleoperated demonstration dataset for training a grasping policy on an SO-ARM101 arm with an AmazingHand dexterous hand (imitation learning).

Task

Pick up the cube with the dexterous hand — grasp a cube on the table using the dexterous hand.

Hardware & Collection

Collection flow: an operator moves the leader arm → the follower tracks it while its gripper proportionally drives the hand's open/close → joint angles and both camera streams are recorded synchronously.

Dataset Statistics

Per-episode flow: start from a fixed initial pose → perform one complete grasp (approach → grasp → lift → move → place) → end.

Position coverage: the cube was placed at 4 different table positions, 5 episodes each, to teach positional generalization.

Data Structure

Joint order: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos

gripper.pos is a virtual dimension — the leader gripper opening, mapped to the hand's grasp pose at training time.

Usage

from lerobot.datasets import LeRobotDataset

ds = LeRobotDataset("Juxi-Technology/soarm_amazing_hand_pick")

Or directly for training:

lerobot-train \
  --dataset.repo_id=Juxi-Technology/soarm_amazing_hand_pick \
  --policy.type=act \
  --policy.device=cuda \
  --steps=60000

Collection Guidelines (Reproduction Notes)

• Consistent initial pose: return the arm pose and object position to the same state before every episode
• Slow, continuous motion: avoid fast swings, which cause large frame-to-frame jumps (especially at 30 fps)
• Reset phases are not recorded: frames captured during reset_time_s are not written to the dataset
• Re-record failures: discard failed attempts and re-record rather than keeping low-quality samples
• Fixed cameras: cameras must not move during collection, and lighting should stay stable

Limitations

• Single scene and lighting condition: training and deployment environments should match
• Limited position coverage: only 4 positions; generalization between them is weaker
• 20 episodes is a small dataset: ACT typically reaches 50–80% success at this scale; 30–50 episodes are recommended for more complex tasks
• No disturbance or failure-recovery samples: the dataset contains only successful demonstrations

Related Model

Trained policy: Juxi-Technology/soarm_amazing_hand_act