Use Execution Modes

Preview: PolicyRuntime and execution modes are planned APIs. The examples below document the target design.

The Execution component decides where inference runs and how requests are scheduled.

Synchronous

This mode runs inference in the runtime thread.

execution:
  class_path: physicalai.runtime.SyncExecution
  init_args:
    mode: chunk

This mode is appropriate for simple deployments and debugging.

Thread Worker

This mode runs inference in a background thread.

execution:
  class_path: physicalai.runtime.AsyncExecution
  init_args:
    transport: thread

Use this mode when model latency should not block robot timing. Since inference backends typically release the GIL, thread-based execution works well for most use cases.

Remote

This mode sends inference requests to a policy server.

execution:
  class_path: physicalai.runtime.RemoteExecution
  init_args:
    endpoint: http://robot-server:8080

Use this mode when the robot host should not hold policy weights or accelerator dependencies.