Installation

Install the runtime package first.

pip install physicalai

Install hardware extras only when you need a specific backend or robot integration.

pip install "physicalai[realsense]"
pip install "physicalai[so101]"
pip install "physicalai[robots]"
pip install "physicalai[capture]"  # all camera backends + IPC transport

For local development, install the repository environment and run the test suite.

uv sync
uv run pytest

Package Boundary

physicalai is the runtime package. It should remain usable on deployment hosts that do not have training dependencies such as Torch or Lightning installed.

Training commands may be provided by a separate training distribution or by plugin entry points.

When installed together with the coordinated physicalai-train release, the shared physicalai executable also exposes training and benchmark subcommands.

Intel GPU runtime (optional)

To run OpenVINO inference on an Intel GPU (device: GPU or AUTO in your inference config), the host needs the OpenCL loader and Intel compute driver. This is the standard Intel GPU host setup on Linux, and these system packages cannot be installed via pip:

# Debian / Ubuntu
sudo apt install ocl-icd-libopencl1 intel-opencl-icd

# allow your user to access the GPU device (log out / back in after this)
sudo usermod -aG render $USER

# verify a GPU device is visible to OpenCL
clinfo -l

Without these, loading the OpenVINO GPU plugin fails with either libOpenCL.so.1: cannot open shared object file (loader missing) or [GPU] Can't get PERFORMANCE_HINT property as no supported devices found (driver missing or device not accessible). CPU inference is unaffected.

Check the Install

import physicalai
from physicalai.inference import InferenceModel

If hardware extras are not installed, importing the base runtime should still work.