Architecture

The Physical AI runtime has a small set of layers with clear boundaries between exported packages, inference, runtime orchestration, and hardware IO.

exported package
    -> InferenceModel
    -> PolicyRuntime
    -> Robot and cameras

Components

Component	Responsibility
Manifest	describes exported artifacts and the inference pipeline
InferenceModel	loads artifacts and computes actions
PolicyRuntime	runs the robot control loop
Execution	decides where inference runs
ActionQueue	buffers and merges action chunks
Robot	reads state and sends commands
Camera	reads image frames

Package Boundary

physicalai is the runtime package. It should not require training dependencies in order to import or run deployment workflows.

Training packages can add commands through CLI entry points when they need to extend the runtime CLI.

physicalai
  infer
  run
  serve

training package
  fit
  validate
  test
  predict
  export

Design Rules

• Config objects remain passive data structures.

• Orchestrators are responsible for executing workflows.

• InferenceModel does not own robot timing.

• PolicyRuntime does not own policy math.

• Manifests describe exported packages.

• Workflow configs describe the desired execution before it starts.