maple.backend.policy.openpi.OpenPIPolicy.act

OpenPIPolicy.act(handle: PolicyHandle, payload: Any, instruction: str, model_kwargs: Dict[str, Any] | None = {}) → List[float]

Get action prediction for a single observation.

Sends visual observations, proprioceptive state, and language instruction to the OpenPI model and receives a predicted action. The model supports multiple observation keys including various camera views and robot state.

OpenPI directly outputs actions in the target action space without requiring explicit unnormalization, similar to SmolVLA.

Parameters:

• handle – Policy handle for the running container.

• payload – Observation payload containing image and state keys. Image keys are automatically detected and base64 encoded. Non-image keys (e.g., ‘state’) are passed through directly.

• instruction – Natural language instruction for the task.

• model_kwargs – Model-specific parameters (optional for OpenPI).

Returns:

Predicted action as list of floats in the target action space.