ContextQMD
Back to Openvino

Inference API Reference

docs/articles_en/physical-ai/reference/inference-api.md

2026.3.01.7 KB
Original Source

Inference API Reference

InferenceModel

python
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

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

from_pretrained

python
model = InferenceModel.from_pretrained(
    "OpenVINO/act-fp16-ov",
    revision="main",      # branch, tag, or commit SHA (optional)
    cache_dir=None,        # download cache directory (optional)
)

This downloads the policy package snapshot from the Hugging Face Hub and loads it like a local export. Additional keyword arguments are forwarded to the constructor (e.g. backend, device).

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

Methods

select_action

python
action = model.select_action(observation)

This method returns one action.

predict_action_chunk

python
chunk = model.predict_action_chunk(observation)

This method returns a chunk of actions for runtime queueing.

reset

python
model.reset()

This method clears the runner state and the action cursor.

close

python
model.close()

This method releases backend resources.

Observation

Observations are dictionaries of NumPy arrays.

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

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