NOJIMA21/so101-egg-transport-act-v0

SO-101 raw egg transport — ACT v0

This is an ACT policy trained for 5,000 steps from 25 human teleoperation demonstrations in the linked LeRobot v3 dataset.

Demonstrated behavior

On the original fixed layout, with no teleoperation during rollout, the policy grasped a raw egg from the pickup area, moved it to an unheated frying pan, and placed it gently. Seven saved autonomous-success rollouts are published with the dataset under evaluations/.

Inputs and outputs

• Two RGB camera views: wrist_camera and overhead_camera
• Six SO-101 joint positions as state
• Six joint-position actions, including the soft gripper
• Recorded and controlled at 30 FPS

Limitations and safety

This is a small fixed-layout proof of a learning loop, not a general-purpose manipulation policy. It was trained with a Hiwonder SO-ARM101 and a soft gripper. Camera placement, calibration, the workspace layout, and the egg's appearance strongly affect behavior. Keep people, hot cookware, and breakable objects out of the workspace, and retain a physical emergency stop when evaluating on hardware.

Reproduction

The associated dataset is available at:

https://huggingface.co/datasets/NOJIMA21/so101-egg-transport-act-v0

Train a new model to a separate output directory; do not overwrite this v0 checkpoint. The policy is published so others can inspect and reproduce the experiment, not as a guarantee of safe transfer to another robot.