fine-tuning-serving-openpi

Fine-tune OpenPI robot policy models by adapting pi0 architectures to custom datasets using JAX or PyTorch backends. The Skill handles model adaptation for ALOHA, DROID, and LIBERO environments, streamlining customization for robot manipulation tasks.

Yes, checkpoint conversion from JAX to PyTorch is supported for deployment. The Skill handles converting JAX checkpoints to PyTorch format, enabling flexible deployment of Physical Intelligence OpenPI models across different frameworks.

OpenPI policy inference supports ALOHA, DROID, and LIBERO environments for robot manipulation tasks. The Skill handles model adaptation and deployment across these specific environments using JAX or PyTorch backends.

No, you can use either JAX or PyTorch to fine-tune and serve OpenPI robot policies. The Skill requires uv>=0.4.0, JAX>=0.4.30 or PyTorch>=2.1.0, along with transformers>=4.53.2 for operation.

Run a policy inference server for real-time control using the serve_policy.py script. After cloning the OpenPI repository and installing dependencies with uv, execute the script with the desired environment parameter like DROID.

Debugging tools for norm stats and GPU memory issues are included with the Skill. These tools help troubleshoot common problems encountered during OpenPI model fine-tuning and policy inference operations.