Runtime package for deploying robot policies trained with Physical AI Studio

Installation • Camera API • Robot API • Inference • Docs

---

Physical AI Runtime provides the deployment-side components for running trained policies on real hardware. It handles camera capture, robot control, and policy inference with a unified API that works across different hardware vendors.

Key Features:

• Unified Camera API — Same interface for UVC, RealSense, Basler, and IP cameras
• Robot Protocol — Structural typing for any robot; no inheritance required
• Inference Engine — Load exported policies from Studio with auto-detected backends
• Policy Runtime — Control loop with observation building and action dispatch

---

Installation

pip install physicalai

With hardware-specific extras:

pip install physicalai[realsense]   # Intel RealSense cameras
pip install physicalai[basler]      # Basler industrial cameras
pip install physicalai[so101]       # SO-101 robot arm
pip install physicalai[trossen]     # Trossen WidowX robots

---

Camera API

All cameras share a unified interface: connect(), read(), read_latest(), and context manager support. Switch hardware without changing application code.

from physicalai.capture import UVCCamera

with UVCCamera(device="/dev/video0", width=640, height=480, fps=30) as camera:
    frame = camera.read_latest()
    print(frame.data.shape)  # (480, 640, 3)
    print(frame.timestamp)   # monotonic timestamp

Intel RealSense (RGB + Depth)

from physicalai.capture import RealSenseCamera

with RealSenseCamera(serial_number="123456789", width=640, height=480, fps=30) as camera:
    rgb, depth = camera.read_rgbd()
    print(rgb.data.shape)    # (480, 640, 3) RGB
    print(depth.data.shape)  # (480, 640) depth in mm

Basler Industrial Camera

from physicalai.capture import BaslerCamera

with BaslerCamera(serial_number="12345678", width=1920, height=1080, fps=60) as camera:
    frame = camera.read_latest()
    print(frame.data.shape)  # (1080, 1920, 3)

Multi-Camera Sync

from physicalai.capture import UVCCamera, RealSenseCamera, read_cameras

cameras = {
    "wrist": UVCCamera(device="/dev/video0"),
    "overhead": RealSenseCamera(serial_number="123456789"),
}

# Connect all
for cam in cameras.values():
    cam.connect()

# Read from all cameras concurrently
synced = read_cameras(cameras)
print(synced.frames["wrist"].data.shape)
print(synced.frames["overhead"].data.shape)

# Cleanup
for cam in cameras.values():
    cam.disconnect()

Camera Discovery

from physicalai.capture import discover_all, UVCCamera

# Discover all connected cameras (returns dict of camera_type -> list of devices)
all_devices = discover_all()
for camera_type, devices in all_devices.items():
    for dev in devices:
        print(f"{camera_type}: {dev.device_id} - {dev.name}")

# Discover specific type
uvc_devices = UVCCamera.discover()

---

Robot API

Robots implement a Protocol-based interface. Any class with connect(), disconnect(), get_observation(), send_action(), and joint_names works — no inheritance required.

from physicalai.robot import SO101

robot = SO101(port="/dev/ttyUSB0")
robot.connect()

obs = robot.get_observation()
print(obs.joint_positions)  # [0.1, 0.2, 0.3, 0.4, 0.5, 0.6]
print(robot.joint_names)    # ['shoulder_pan', 'shoulder_lift', ...]

robot.send_action(target_positions, goal_time=0.1)
robot.disconnect()

Trossen WidowX-AI

from physicalai.robot import WidowXAI

robot = WidowXAI()
robot.connect()

obs = robot.get_observation()
print(obs.joint_positions)

robot.send_action(target_positions)
robot.disconnect()

Bimanual WidowX-AI

from physicalai.robot import BimanualWidowXAI

robot = BimanualWidowXAI()
robot.connect()

obs = robot.get_observation()
# Joint positions for both arms concatenated
print(obs.joint_positions.shape)

robot.send_action(bimanual_targets)
robot.disconnect()

Robot Verification

from physicalai.robot import SO101, verify_robot

robot = SO101(port="/dev/ttyUSB0")
verify_robot(robot)  # Interactive joint-by-joint check

---

Inference

Load exported policies from Physical AI Studio. The InferenceModel class auto-detects the backend (OpenVINO or ONNX in this package; companion distributions may contribute additional adapters such as ExecuTorch) and handles action chunking automatically.

from physicalai.inference import InferenceModel

# Load exported policy
model = InferenceModel.load("./exports/act_policy")

# Reset state for new episode
model.reset()

# Run inference
action = model.select_action(observation)

With Explicit Backend

from physicalai.inference import InferenceModel

# Force specific backend
model = InferenceModel.load(
    "./exports/act_policy",
    backend="openvino",
    device="GPU",
)

Latency benchmarking

import json

from physicalai.benchmark.performance import InferenceLatencyBenchmark
from physicalai.inference import InferenceModel

model = InferenceModel.load("./exports/act_policy")
model.reset()
benchmark = InferenceLatencyBenchmark(
        max_iters=100,
        warmup_iters=2,
        max_duration=10000,
    )
metrics = benchmark.run(model)
print(json.dumps(metrics, indent=2))

---

Policy Runtime

The PolicyRuntime orchestrates the full control loop: connecting hardware, reading cameras, building observations, running inference, and dispatching actions to the robot.

from physicalai.runtime import PolicyRuntime, SyncExecution
from physicalai.inference import InferenceModel
from physicalai.capture import UVCCamera, RealSenseCamera
from physicalai.robot import SO101

runtime = PolicyRuntime(
    fps=30,
    robot=SO101(port="/dev/ttyACM0"),
    model=InferenceModel.load("./exports/act_policy"),
    cameras={
        "wrist": UVCCamera(device="/dev/video0", width=640, height=480),
        "overhead": RealSenseCamera(serial_number="123456789"),
    },
    execution=SyncExecution(),
)

with runtime:
    runtime.run(duration_s=60)

From YAML Config

Preview: This API is not yet implemented.

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

# runtime.yaml
runtime:
  class_path: physicalai.runtime.PolicyRuntime
  init_args:
    fps: 30
    robot:
      class_path: physicalai.robot.so101.SO101
      init_args:
        port: /dev/ttyACM0
    model:
      class_path: physicalai.inference.InferenceModel
      init_args:
        export_dir: ./exports/act_policy
    cameras:
      wrist:
        class_path: physicalai.capture.UVCCamera
        init_args:
          device: /dev/video0
          width: 640
          height: 480
    execution:
      class_path: physicalai.runtime.SyncExecution
      init_args:
        mode: chunk

CLI

physicalai run --config runtime.yaml --run.duration_s=60

The runtime package owns the shared physicalai executable. Training packages can add subcommands such as fit and benchmark through the physicalai.cli.subcommands entry-point group.

Async Execution

Async execution runs inference in a background thread while the main loop handles camera reads and robot commands at a fixed frequency. Useful when inference is slower than the control rate.

from physicalai.runtime import PolicyRuntime, AsyncExecution

runtime = PolicyRuntime(
    fps=30,
    robot=robot,
    model=model,
    cameras=cameras,
    execution=AsyncExecution(fps=30),
)

with runtime:
    runtime.run(duration_s=60)

Remote Execution

Remote execution sends observations to an inference server and receives actions over the network. Useful for running large models on a separate GPU machine.

Preview: This API is not yet implemented.

from physicalai.runtime import PolicyRuntime, RemoteExecution

runtime = PolicyRuntime(
    fps=30,
    robot=robot,
    cameras=cameras,
    execution=RemoteExecution(endpoint="http://gpu-server:8080/infer"),
)

runtime.run(duration_s=60)

---

Full walkthrough: See examples/tutorials/collect_train_deploy.ipynb for a complete collect -> train -> deploy guide.

---

Documentation

Home • Getting Started • How-To Guides • Concepts • API Reference

Contributing

See CONTRIBUTING.md.