Instantiate Components

A Config recipe points at one class and the arguments for its constructor.

class_path: physicalai.capture.UVCCamera
init_args:
  device: /dev/video0
  width: 640
  height: 480

Build from a recipe

Config.instantiate() creates a new object from a trusted local recipe. It runs __init__ only—you still call connect(), run(), or similar yourself.

from physicalai.config import Config

config = Config.from_dict({
    "class_path": "physicalai.capture.UVCCamera",
    "init_args": {"device": "/dev/video0", "width": 640, "height": 480},
})
camera = config.instantiate()
camera.connect()

Save how an object was created

Add @export_config to a class so Physical AI can record the constructor arguments you actually passed (omitted defaults stay omitted).

from physicalai.capture import UVCCamera
from physicalai.config import Config

camera = UVCCamera(device="/dev/video0", width=640, height=480)
config = Config.from_instance(camera)

Nested components use the same recipe shape inside init_args. Values must be JSON-friendly: paths become strings, tuples become lists, and invalid floats are rejected.

Config.from_instance(camera).save("camera.yaml")

Safety

class_path loads and runs Python code on your machine. Use your own config files or other sources you trust. Skip instantiation for metadata or messages that came from another process or the network.

Inference manifests (policy exports) use a separate format with aliases and artifact paths. Use Config for runtime construction YAML; use manifest APIs when loading an exported policy package.