PhysicalAI OpenArm Plugin

Direct SocketCAN/Damiao control for OpenArm followers and hand-guided leaders in PhysicalAI. The plugin also registers single-arm and bimanual OpenArm entries in Physical AI Studio.

Safety and support

OpenArm moves powerful physical hardware. Secure each arm, provide clearance, keep an emergency stop accessible, and follow OpenArm's official safety guide. Only Linux SocketCAN is supported. Exactly one process may control a CAN interface at a time: do not concurrently run this plugin with LeRobot, ROS 2, or other CAN clients on the same bus.

This first release assumes hardware setup has already been completed. It does not set motor IDs, write persistent parameters, calibrate, or zero encoders. Before connecting, configure the documented motor IDs, provision CAN/CAN-FD, and complete OpenArm's homing and closed-gripper initialization procedure.

Installation

uv add physicalai-openarm-plugin

The package requires python-can and Linux SocketCAN. OpenArm's default configuration uses CAN-FD at 1 Mbps arbitration and 5 Mbps data rate.

CAN adapters

can_adapter: socketcan is the default and supported route. Use a Linux SocketCAN channel such as can0; a USB-CAN adapter is supported when the host exposes it as a SocketCAN interface. This includes adapters configured through slcand, which usually appear as slcan0 rather than /dev/ttyUSB0.

An experimental can_adapter: damiao proof of concept uses motorbridge to open a Damiao USB-CAN serial adapter directly. Set port to its device path and optionally set dm_serial_baud (default 921600):

port: /dev/ttyACM0
can_adapter: damiao
dm_serial_baud: 921600

This serial path has not been tested with OpenArm's eight-motor topology or CAN-FD timing. Do not use it for unattended motion, and retain SocketCAN as the recommended path until an adapter-specific hardware acceptance test succeeds.

Runtime contract

Positions are degrees, in this fixed order:

joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, joint_7, gripper

Followers require side: left or side: right; the plugin applies the side-specific conservative limits used by LeRobot's OpenArm support. Leaders are read-only and default to torque-disabled manual control. Native bilateral force feedback is not implemented.

Run

Update the can0 and can1 placeholders, then run unilateral teleoperation:

uv run physicalai run --config packages/physicalai-openarm-plugin/examples/runtime/teleop.yaml

For two leaders and two followers on independent CAN interfaces:

uv run physicalai run --config packages/physicalai-openarm-plugin/examples/runtime/bimanual-teleop.yaml

To read and print a leader's joint positions without sending CAN position targets:

uv run physicalai run --config packages/physicalai-openarm-plugin/examples/runtime/read-joints.yaml

Studio assets

The plugin includes lightweight parallel-link gripper URDF visualization models for Studio. They are plugin-authored kinematic visualizations, not redistributed OpenArm CAD or mesh assets. OpenArm software is Apache-2.0; hardware design data has separate CERN-OHL-S-2.0 licensing.

Development

uv sync
uv run pytest packages/physicalai-openarm-plugin/tests/