Add Runtime Callbacks

Preview: PolicyRuntime and runtime callbacks are planned APIs. The examples below document the target design.

Use callbacks when you need product-specific behavior around the runtime loop.

The following example records observations and actions.

python

class RecordingCallback:
    def on_observation(self, observation, step):
        recorder.write_observation(step.t, observation)

    def before_send_action(self, action, step):
        recorder.write_policy_action(step.t, action)
        return action

    def on_action_sent(self, action, step):
        recorder.write_sent_action(step.t, action)

    def on_stop(self):
        recorder.close()

Attach the callback when you construct the runtime.

python

runtime = PolicyRuntime(
    robot=robot,
    model=model,
    execution=execution,
    fps=30,
    callbacks=[RecordingCallback()],
)

As a general rule, keep workflow-specific logic in callbacks unless the same behavior becomes a reusable runtime primitive.