maple.backend.policy.openpi.OpenPIPolicy
- class maple.backend.policy.openpi.OpenPIPolicy
Backend for OpenPI (π₀ / π₀.₅) vision-language-action models.
OpenPI is Physical Intelligence’s family of generalist robot policies that condition on visual observations, proprioceptive state, and natural language instructions to predict robot actions. The models support multiple observation modalities and are available in different sizes (π₀ and π₀.₅) and variants (base, fast, task-specific fine-tunes).
The backend manages Docker containers running the OpenPI inference server, downloads model weights from public S3 storage using fsspec with anonymous access, and serves predictions via HTTP API.
Methods
act(handle, payload, instruction[, model_kwargs])Get action prediction for a single observation.
info()Get policy backend information and capabilities.
pull(version, dst)Pull model weights and Docker image.
pull_gs(version, dst)Pull model weights from Google Cloud Storage and Docker image.
serve(version, model_path, device[, ...])Start serving OpenPI model in a Docker container.
Attributes