SO-101 AUTONOMOUS ARM

An end-to-end physical AI build on the open-source SO-101 (SO-ARM100) 5-DOF arm with parallel gripper. Demonstrations were collected by teleoperating a twin leader arm; an ACT (Action Chunking Transformer) policy was trained with Hugging Face LeRobot on a rented RTX 4090 and deployed back to the real arm, which completed the pick-and-place autonomously.

The first trained policy lifted the arm 5 mm and froze: 34.5% of the recorded frames were static dwell at the start and end of each demo, and ACT's 100-action chunks turned that into a strong 'do nothing' prior. An automated kinematic trimming pipeline cut the dataset from 21,665 frames to 14,201 pure-motion frames, and the retrained policy completed the task.

Windows + WSL2 is a trap for real-time robotics: usbipd camera passthrough starved the 30 Hz control loop down to 4.6 Hz, so the camera runs through a native Windows DirectShow MJPEG server consumed by WSL over HTTP. Cloud training was wrapped in triple budget watchdogs (local cost cap, in-container watchdog, guaranteed cleanup) and finished at $2.23 for 47,700 steps. Physical rollout debugging covered leader/follower serial-port calibration mixups, a policy gone blind in a dark room (mean pixel brightness 4.33), and out-of-distribution starting poses.