fine-tuning-serving-openpi - Fine-Tune and Serve OpenPI Robot Policies
Fine-tunes and serves Physical Intelligence OpenPI models with JAX or PyTorch for robot policy inference in ALOHA, DROID, and LIBERO environments.
Tags
Updated: 2026-09-28Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Fine-tune OpenPI models
- Serve policy inference
- Convert JAX checkpoints
- Compute normalization statistics
- Debug training issues
Inputs
- OpenPI repository
- Training dataset
- Model checkpoint
- Training configuration
- Target environment
- Robot observations
Outputs
- Trained model checkpoints
- Converted PyTorch checkpoints
- WebSocket inference server
- Policy action arrays
- Normalization statistics
Requirements
- uv 0.4 or later
- JAX 0.4.30 or later
- PyTorch 2.1 or later
- Transformers 4.53.2 or later
- OpenPI repository with submodules
- GPU resources for training or serving