fine-tuning-serving-openpi

To fine-tune OpenPI policies, prepare your config and compute normalization statistics for your custom dataset. You can train pi0, pi0-fast, or pi0.5 policies using ALOHA, DROID, or LIBERO datasets.

You can convert JAX checkpoints to PyTorch format to ensure compatibility during policy serving. This conversion step is required before launching the WebSocket inference server for robot control.

OpenPI fine-tuning fails because normalization statistics are required to process custom datasets correctly. You must compute and provide these statistics in your config before training pi0, pi0-fast, or pi0.5 policies.

Yes, you can launch a WebSocket policy server for real-time robot inference. After converting your JAX checkpoints to PyTorch format, the server handles inference for your trained pi0, pi0-fast, or pi0.5 policies.

You need optional transformer patching to resolve PyTorch compatibility issues when converting OpenPI checkpoints. The built-in troubleshooting guidance also helps handle config mismatches and backend memory limits.

OpenPI fine-tuning works directly with DROID and LIBERO datasets. The workflow applies to pi0, pi0-fast, and pi0.5 policies, allowing you to prepare configs, train, and serve models for these specific robotics datasets.