Runtime

Preview: PolicyRuntime is a planned API. The content below documents the target design.

PolicyRuntime runs a policy on robot hardware. It owns the control loop, the callback lifecycle, and the interaction between observations, inference requests, and actions.

runtime = PolicyRuntime.from_config("runtime.yaml")
runtime.run(duration_s=60)

Responsibilities

Component	Owns	Does not own
InferenceModel	model load, preprocess, inference, postprocess	robot loop timing
Execution	where inference runs	robot IO
ActionQueue	action chunks and buffering	model inference
PolicyRuntime	observe, request inference, send action, callbacks, timing	policy math
Robot	hardware connection, observations, actions	policy inference

Loop

The runtime loop follows this general pattern:

while running:
    observation = get_robot_state() + get_camera_frames()
    maybe_request_inference(observation)
    action = get_next_action_or_hold()
    send_action_to_robot(action)
    sleep_until_next_tick()

The exact observation structure and merging strategy may change as the API stabilizes.

Execution Modes

Mode	Where inference runs	Use
SyncExecution(mode="single_action")	runtime thread	simple policies
SyncExecution(mode="chunk")	runtime thread	chunk policies without background worker
AsyncExecution(transport="thread")	worker thread	avoid blocking control loop
RemoteExecution	remote server	robot host without policy weights

Product Workflows

HIL, recording, highlight, and DAgger should be composed through callbacks until they justify reusable runtime primitives.

class HILCallback:
    def before_send_action(self, action, step):
        if teleop.enabled:
            return teleop.read_action()
        return action