What problem does it solve? Adapting Physical Intelligence's OpenPI vision-language-action models (pi0, pi0-fast, pi0.5) to custom robot datasets and deploying them as low-latency inference servers involves complex setup, checkpoint conversion, and debugging that this Skill streamlines into repeatable workflows. ## Core Features & Use Cases - JAX and PyTorch Fine-Tuning: Config-driven training on LeRobot or RLDS datasets with normalization statistics computation and multi-GPU FSDP/DDP support. - Checkpoint Conversion: Convert JAX checkpoints to PyTorch format with required transformers patches for deployment-friendly inference. - Policy Serving: Launch WebSocket policy servers for ALOHA, DROID, and LIBERO environments with remote client integration patterns. - Use Case: A robotics engineer fine-tunes pi0.5 on a custom DROID dataset, computes norm stats, trains on an A100 cluster, then serves the checkpoint so a robot client can request action chunks over WebSocket. ## Quick Start Ask the agent to fine-tune an OpenPI pi0.5 model on your LeRobot dataset and serve the resulting checkpoint as a policy inference server.
