act-okura-kinesthetic-wrist-7d

ACT policy for the okra grasp on a Unitree G1, trained on kinesthetic teaching data. Takes over from an IK pre-grasp reach to do the final approach.

I/O

• input: observation.state (7,) = right-arm joint angles (shoulder p/r/y, elbow, wrist r/p/y; G1 motors 22-28) + one camera observation.images.cam_right_wrist (480x640). The head camera is NOT used (during kinesthetic teaching it sees the operator's arm = OOD; head is used only for the IK target click).
• output: action (7,) absolute right-arm joint angles, ACT chunk_size 100.
• gripper is NOT part of this policy (7-DoF arm only). The Dex1-1 open/close (and the eventual cut-and-grasp end-effector) is handled outside ACT.

Training

• dataset: sotata/okura-kinesthetic-wrist-7d (10 episodes / 652 frames / 30fps slice, LeRobot v3.0).
• policy: ACT (lerobot), batch 8, chunk 100, steps 100000.
• GPU: AWS g6e.2xlarge (L40S 48GB), ~0.045 s/step, ~1h35m.
• final loss ~0.005 (small dataset -> strongly overfit; validate on hardware).
• W&B: lerobot-act-okura / run w5wxoffl.

Load

--policy.path=sotata/act-okura-kinesthetic-wrist-7d

Deploy note

Pipeline: IK (click -> right-arm pre-grasp) -> handoff -> this ACT (wrist-only, 7-DoF) for the final approach. The inference side must be wired for 7-DoF state/action + the single wrist camera (differs from the prior 2-camera / 8-DoF okura-pick-tree models).