What problem does it solve? Trained robot-control policies (LeRobot ACT, Pi0/openpi VLA) cannot run directly on RDK boards — the BPU only executes OE-compiled .hbm files. This Skill guides the full export → compile → on-board control loop so a trained checkpoint actually drives a robot arm on RDK S100/S600. ## Core Features & Use Cases - ACT deployment (three stages): Export a trained ACT checkpoint to two ONNX submodels with export_bpu_actpolicy.py, compile to .hbm via build_all.sh inside the OE 3.7.0 Docker toolchain, then run bpu_control_robot.py on the board. - Branch/board routing: Picks the correct rdk_LeRobot_tools branch (s100/s600/stable), march (nash-e vs nash-p), and LeRobot version (upstream v0.5.2 vs the legacy D-Robotics fork). - Pi0/openpi VLA on S600: Sets up the client-server runtime — s600_inference_node + piper_node on the board, Pi0 inference on the OE-LLM server (port 8888) — with the correct HBM model and norm_stats.json. - Use Case: You trained an ACT policy on an SO-101 arm and want it running on an RDK S100: the Skill walks you through ONNX export with type: nash-e, OE Docker compilation, building the Python-3.12 bpu_runtime C++ extension, and launching the control loop. ## Quick Start Ask the agent to deploy your trained ACT policy to an RDK S100, providing the checkpoint path and target board, and it will walk you through ONNX export, .hbm compilation, and the on-board control loop.
