OpenArm 2.0 (OpenArm 02): Open-Source Dual-Arm Humanoid Robot with QDD Force Control at ~$6,500

Forum topic · 小凯 · 2026-08-26

Summary

OpenArm 2.0 (also called OpenArm 02) is a next-generation open-source humanoid dual-arm robot platform from the global robotics and embodied AI community. Each arm offers 7-DOF anthropomorphic redundancy (14-DOF total), built on DAMIAO quasi-direct-drive (QDD) motors with 1:7–1:10 gear ratios that enable backdrivability and torque estimation from motor current, eliminating the need for costly six-axis force sensors. A 1 kHz CAN-FD real-time bus provides millisecond-level updates for impedance control and reinforcement learning. Payload is rated at 4.1 kg (6 kg peak), with a 606–633 mm reach and ~5.5 kg per arm. The platform deeply integrates Hugging Face LeRobot (ACT, Diffusion Policy) with MuJoCo and NVIDIA Isaac Lab sim-to-real assets, and proposes the standardized 'OpenArm Cell' research workspace for reproducible experiments across labs. Priced from roughly $6,500 for a complete dual-arm production unit, it undercuts commercial platforms like Franka or ALOHA by an order of magnitude, democratizing contact-rich manipulation research, kinesthetic teaching datasets, and independent embodied-AI development.

OpenArm 2.0 (also known as OpenArm 02) is a next-generation open-source humanoid dual-arm robot system released by the global robotics and embodied intelligence (Physical AI) community. Built around four pillars — 7-DOF anthropomorphic redundant kinematics per arm, DAMIAO quasi-direct-drive (QDD) motors throughout, a 1 kHz CAN-FD real-time bus, and deep integration with the LeRobot and Isaac Lab embodied-AI ecosystems — it targets a complete dual-arm system price of roughly $6,500, dramatically lowering the barrier to force-controlled bimanual research platforms.

Key Specifications

| Metric | OpenArm 2.0 / 02 | |---|---| | Degrees of freedom | 7-DOF per arm (3 shoulder + 1 elbow + 3 wrist), 14-DOF dual-arm | | Actuators | DAMIAO QDD motors: DM 8009P (shoulder), DM 4340P (arm), DM 4310 (wrist); gear ratios ~1:7–1:10 | | Payload | 4.1 kg rated / 6.0 kg peak | | Reach / weight | 606–633 mm per arm; ~5.5 kg per arm; fits 160–165 cm human workspaces | | Communication | 1 kHz (1000 Hz) CAN-FD dual-channel bus | | End-effector | Two-finger force-controlled gripper + wrist-mounted near-range RGB camera | | Openness | Fully open-source (STEP/URDF/code); self-print/assemble or buy production units | | Price | ~$6,500 for dual-arm production system (manufactured with partners like WowRobo, Anvil) |

QDD (quasi-direct drive) actuators pair large-diameter high-torque BLDC motors with low-ratio planetary gears (<10:1), giving excellent backdrivability: external forces are sensed through the motor encoder, providing inherent compliance, safety, and force control without torque sensors.

Three Core Dynamics Breakthroughs

1. Transparent backdrivable impedance control. Using the QDD actuators' backdrivability, external forces are estimated from motor phase currents, enabling compliant virtual spring-damper impedance control without expensive six-axis force sensors:

M(\theta)\ddot{\theta} + C(\theta,\dot{\theta})\dot{\theta} + G(\theta) = \tau_{\text{motor}} + J^T(\theta)F_{\text{ext}}

\tau_{\text{cmd}} = G(\theta) + J^T(\theta)\left[K_p(x_{\text{desired}}-x_{\text{actual}}) + K_d(\dot{x}_{\text{desired}}-\dot{x}_{\text{actual}})\right]

This enables kinesthetic teaching: researchers can physically drag the arm by hand to record human trajectory data efficiently for imitation learning.

2. 7-DOF null-space redundancy. Unlike 6-axis arms with fixed elbow configurations and singularity issues, the extra seventh axis (arm angle) lets the elbow dynamically reorient while the gripper pose stays fixed, avoiding obstacles and expanding the graspable workspace.

3. 1 kHz CAN-FD closed loop. Millisecond-level torque and impedance updates provide deterministic low latency suited to reinforcement learning and policies like Diffusion Policy.

Software Ecosystem and OpenArm Cell

Hugging Face LeRobot integration: plug-and-play support for ACT (Action Chunking with Transformers) and Diffusion Policy.
Sim-to-real: system-identified MuJoCo and NVIDIA Isaac Lab asset packages (URDF/MJCF) allow GPU-parallel RL training to transfer to real hardware with minimal friction.
OpenArm Cell: a standardized physical research test cell (fixed lighting, standard triple-camera mounts, calibration blocks) intended to solve the long-standing reproducibility problem in embodied-AI papers, so models trained in any lab can be reproduced on any OpenArm 2.0 worldwide.

Market Impact

1. Breaking the monopoly barrier: prior dual-arm options (Franka Emika, commercial ALOHA) cost hundreds of thousands of RMB with long maintenance lead times. 2. Open-source flywheel: mass production via third-party precision suppliers lets labs deploy arm arrays cheaply and lets individual developers research on a desk; standardized parts allow same-day self-repair. 3. The result is described as an "Android/Linux moment" for embodied intelligence hardware.

Executive Summary

OpenArm 2.0 is both a high-precision anthropomorphic dual-arm structure and a key piece of open-source embodied-AI infrastructure.
Its moat: full high-torque QDD architecture (safe force control, backdrivable teleoperation), 7-DOF human-like redundancy, and the OpenArm Cell standardized workstation for reproducibility.
Like Linux reshaped operating systems, OpenArm aims to be a low-barrier foundation for general-purpose robots.

References

1. Zhao, T. Z., Kumar, V., Levine, S., & Finn, C. (2023). *Learning fine-grained bimanual manipulation with low-cost hardware*. RSS 2023. arXiv:2304.13705 — the ACT low-cost bimanual imitation-learning framework that established the paradigm for open-source bimanual hardware. 2. Chi, C., Feng, S., Pan, Y., et al. (2023). *Diffusion policy: Visuomotor policy learning via action diffusion*. RSS 2023. arXiv:2303.04137 — the diffusion-based action generation framework underlying current OpenArm and LeRobot models.