maple.backend.policy.openvla.OpenVLAPolicy.act

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

Get action prediction for a single observation.

Sends a visual observation and language instruction to the OpenVLA model and receives a predicted action. The action is unnormalized using dataset statistics specified by unnorm_key.

IMPORTANT: OpenVLA requires unnorm_key to be specified. The model outputs normalized actions that must be converted to the target action space using dataset-specific statistics. Without unnormalization, the actions cannot be executed on real robots or simulators.

Parameters:

• handle – Policy handle for the running container.

• payload – Observation payload containing ‘image’ key with image data.

• instruction – Natural language instruction for the task.

• model_kwargs – Model-specific parameters. Must contain ‘unnorm_key’. (REQUIRED). Examples: ‘libero_spatial’, ‘bridge’, ‘fractal’.

Returns:

Predicted action as list of floats, unnormalized to target space.