NOJIMA21/so101-egg-transport-act-v0

SO-101 raw egg transport: teleoperation and first ACT rollout

This dataset documents a single-arm Hiwonder SO-ARM101 task: pick up a raw egg from a fixed pickup area and gently place it in an unheated frying pan.

Contents

• Root dataset: 25 human teleoperation demonstrations in LeRobot v3 format.
• evaluations/autonomous_success_001/ through evaluations/autonomous_success_007/: seven saved autonomous-success rollouts of an ACT policy trained from the 25 demonstrations.

Training data

The root is directly loadable as a LeRobot dataset. It contains two camera views (wrist_camera, overhead_camera), 6-DoF joint state, and 6-DoF action, recorded at 30 FPS. The soft gripper has no force sensor.

Task instruction:

Pick up the raw egg from the pickup area and gently place it in the unheated frying pan without cracking it.

Autonomous evaluation

The included evaluations were produced with ACT, trained for 5,000 steps from the 25 root demonstrations. They are seven saved autonomous raw-egg grasp and placement cycles with no teleoperation. Each evaluation contains camera video, state, and policy-produced actions. evaluation.json records the exact checkpoint and outcome for the first independently annotated rollout; the other six are preserved as the raw LeRobot evaluation records.

After the episode had been finalized, the follower bus reported a Torque_Enable write failure for motor ID 3 while disconnecting. This communication issue did not prevent the episode from being saved, and is retained as an operational limitation rather than hidden.

Safety and limitations

This is a small, fixed-layout dataset. It is a demonstration of an initial learning loop, not evidence of robust general-purpose manipulation. Do not run the policy around people, hot cookware, or breakable objects without independent safety checks.