vllm_omni.entrypoints.openpi.serving ¶

Serving layer for robot policy inference via /v1/realtime/robot/openpi.

Flow: raw obs → engine request → actions. The loaded policy model owns dataset transforms inside its pipeline.

ActionOutput `module-attribute` ¶

ActionOutput = ndarray | dict[str, ndarray]

logger = init_logger(__name__)

OpenPI policy server handshake config.

Values are model-specific and must be provided by the loaded policy model.

values: dict[str, Any]

from_model_config(model_config: Any) -> PolicyServerConfig

to_dict() -> dict[str, Any]

Robot policy serving layer for OpenPI protocol.

Model-specific transform/state lives in the diffusion pipeline.

engine_client = engine_client

model_name = model_name

policy_server_config = _get_policy_server_config(
    engine_client
)

create_policy_server(
    engine_client: Any, model_name: str | None = None
) -> ServingRealtimeRobotOpenPI | None

infer(
    obs: dict, *, session_id: str, reset: bool
) -> ActionOutput

raw obs → engine → actions.

reset(obs: dict) -> None

Compatibility hook; per-connection state lives in RobotRealtimeConnection.