maple.backend.policy.openpi

OpenPI policy backend.

This module implements the policy backend for OpenPI (π₀ / π₀.₅), Physical Intelligence’s vision-language-action models for robotic manipulation. OpenPI takes visual observations, proprioceptive state, and natural language instructions as input and outputs robot actions.

OpenPI models are downloaded from the public S3 bucket gs://openpi-assets using fsspec with anonymous access (no credentials required). The models support both base versions for fine-tuning and pre-trained versions for specific datasets including DROID, ALOHA, and LIBERO.

Available versions: - Base models: pi0_base, pi0_fast_base, pi05_base (for fine-tuning) - DROID: pi0_fast_droid, pi0_droid, pi05_droid (mobile manipulation) - ALOHA: pi0_aloha_towel, pi0_aloha_tupperware, pi0_aloha_pen_uncap (bimanual tasks) - LIBERO: pi05_libero (long-horizon manipulation benchmark) - latest: Alias for pi05_droid (default)

Classes

OpenPIPolicy()

Backend for OpenPI (π₀ / π₀.₅) vision-language-action models.