OpenARM Quick Start Guide: An Open-Source Backdrivable Dual-Arm Robot for Physical AI

Key points

• What OpenARM is: A fully open-source, 7-DOF-per-arm dual-arm robot with quasi-direct-drive (QDD) joints, designed for physical AI and embodied intelligence research. Hardware CAD, firmware, control code, and simulation tools are all public.
• Defining features: Full backdrivability (you can push the arm by hand and it yields) and bilateral force feedback (the follower arm returns contact forces to the operator's leader arm), enabling safe hand-guided teaching and tactile teleoperation.
• Three-piece ecosystem:
• OpenArm 2.0 – the arm itself: 633 mm reach, 4.1 kg rated / 6.0 kg peak payload, QDD backdrivable joints with mechanical limits per axis, parallel gripper with integrated hand-eye camera, MISUMI aluminum frame.
• OpenArm Cell – a standardized reproducible evaluation chamber (uniform background, lighting, cameras, arm placement) for fair, automated model comparison.
• OpenArm KER – Kinematic Equivalent Replica, a motor-less leader arm matching the real arm's kinematics for low-fatigue teaching.
• Core concepts explained: 7 DOF, QDD quasi-direct-drive, CAN-FD bus (1 Mbps, 1 kHz control loop), ROS2 (ros2_control + MoveIt2), unilateral vs bilateral teleoperation, and LeRobot (Hugging Face's imitation-learning framework, which officially uses OpenARM as a data-collection device).
• Prerequisites: Ubuntu 22.04 or 24.04, ROS2 Humble/Iron/Jazzy, a USB-CAN adapter, Python 3.10+ in a venv, colcon + ament build tools. The Damiao variant specifically requires Ubuntu 24.04, a candleLight-firmware USB-CAN, and a pure Python venv workflow.
• Seven-step quickstart (enactic mainline):
1. Install ROS2 and base dependencies (python3-colcon-common-extensions, can-utils). 2. Create ~/ros2_ws/src and clone seven repos: openarm_description, openarm_can, openarm_ros2, openarm_teleop, and optionally openarm_isaac_lab and openarm_mujoco. 3. colcon build and source the install. 4. Bring up SocketCAN: sudo ip link set can0 up type can bitrate 1000000. 5. Visualize in RViz with display_openarm.launch.py (single or bimanual, arm_type:=v10). 6. Launch bringup in fake-hardware mode and publish a JointState to move to the home position; start with small, slow motions to avoid self-collision. 7. Run unilateral keyboard teleop (teleop_unilateral.launch.py) or bilateral force-feedback teleop (teleop_bilateral.launch.py); for real hardware pass can_interface:=can0.
• Advanced paths:
• Sim2Real: train PPO and other RL policies in parallel on GPU inside openarm_isaac_lab, then transfer to the real arm.
• Imitation learning: record human demonstrations via bilateral teleop as LeRobot datasets and train policy networks.
• Reproducible evaluation: place the arm in the OpenArm Cell so model A vs model B comparisons are automatable and rerunnable.
• Two code branches to pick by hardware:
• enactic/openarm mainline – ROS2 + CAN-FD + Damiao DM-series motors (J8009P/J4340P/J4310); follow the seven steps above.
• anvil-robotics/openarm (Damiao variant) – Ubuntu 24.04, pure Python venv, requires candleLight-firmware USB-CAN; setup uses scripts/setup_can.sh to assign persistent CAN names, then pip install -e . and runs python -m openarm.damiao.gravity for gravity-compensation demos and python -m openarm.damiao.monitor -t for leader–follower teleoperation. A separate Dynamixel (6-DOF) community guide covers single-arm RViz + keyboard teleop via display.launch.py and keyboard_teleop.
• Safety: backdrivability means the arm can pinch or jerk unexpectedly. Bolt it securely to a stable bench away from foot traffic and flammables, know the e-stop location, wear PPE (safety glasses, fitted sleeves), keep payloads below 4.1 kg after subtracting gripper weight, and re-assess joint limits for large or unusual end-effectors.
• Resources:
• Home: https://openarm.dev
• Docs: https://docs.openarm.dev
• Code: https://github.com/enactic/openarm
• Simulation: openarm_isaac_lab
• Dataset: openarm_dataset
• Dataflow: dora-openarm
• Community: Discord https://discord.gg/GmYa262ETH and GitHub Issues/Discussions
• Procurement: DIY kits or assembled units; free manufacturing and derivative-development license granted to Chinese mainland manufacturers.

Bottom line: OpenARM is not just "a robot arm" but an open, reproducible, tactile training and benchmarking arena for robotic manipulation. By combining full backdrivability, bilateral force feedback, and a standardized evaluation cell, it turns "my model is better than yours" into evidence that can be rerun and compared under identical conditions.

*Specifications and commands compiled from openarm.dev, docs.openarm.dev, the enactic and anvil-robotics repositories, and community onboarding materials. The project iterates quickly; always defer to the documentation matching your specific hardware revision.*