End-to-End Imitation Learning for SO-101 with ROS 2

Developer Dmitri Manajev has launched SO-101 ROS Physical AI, an open-source ROS 2 stack designed specifically for the affordable SO-101 robot arm. The project creates a complete, practical workflow for imitation learning on real hardware, addressing a gap in available resources. It includes everything from robot bringup and ros2_control integration to leader/follower teleoperation for collecting demonstrations and multi-camera support for episode recording. The system converts rosbag/MCAP files into LeRobot datasets and supports popular policies including ACT, SmolVLA, and π0/π0.5.

The architecture enables both synchronous on-device inference and asynchronous remote inference via a policy server, allowing computationally intensive models to run on separate GPU hardware while the robot executes actions locally. This makes advanced imitation learning accessible even with limited onboard compute. The typical workflow involves teleoperating the robot to collect demonstrations, converting recordings to LeRobot format, training policies, and deploying them back to the physical robot through ROS 2. Manajev designed the project specifically for students, researchers, and hobbyists who want to learn learning-based control in robotics with practical, reusable examples rather than isolated demos.

Demo videos showcase key components including async policy inference with remote servers, the episode recording workflow, and teleoperation setup. The project bridges the ROS 2 and robot learning ecosystems by providing a single reference implementation covering the entire imitation learning pipeline on real hardware. As an actively developed open-source project, it serves both as a practical platform for experimentation and as an educational resource for those exploring modern robotics techniques.