Inference API Reference

InferenceModel

InferenceModel(
    export_dir: str | Path,
    policy_name: str | None = None,
    backend: str = "auto",
    device: str = "auto",
    runner: InferenceRunner | None = None,
    preprocessors: list[Preprocessor] | None = None,
    postprocessors: list[Postprocessor] | None = None,
    callbacks: list[Callback] | None = None,
    **adapter_kwargs,
)

The model can be constructed directly from an export directory or loaded from config.

Constructors

model = InferenceModel.load("./exports/act_policy")

Note: InferenceModel.from_config() is a planned API.

Methods

select_action

action = model.select_action(observation)

This method returns one action.

predict_action_chunk

chunk = model.predict_action_chunk(observation)

This method returns a chunk of actions for runtime queueing.

reset

model.reset()

This method clears the runner state and the action cursor.

close

model.close()

This method releases backend resources.

Observation

Observations are dictionaries of NumPy arrays.

observation = {
    "state": joint_positions,
    "image.wrist": wrist_image,
}

The expected keys and shapes come from the exported package and its preprocessing pipeline.