OpenArm 2.0 Standardizes Cameras and Lighting with a Cell to Address the Cross-Lab Evaluation Gap in Physical AI

OpenArm updated its 2.0 documentation on September 14, expanding the project from a standalone open-source robotic arm into a physical-AI platform spanning hardware, demonstration, datasets, and evaluation environments. The core arm retains its human-scale 7 degrees of freedom, with a rated payload of 4.1 kg and a peak payload of 6.0 kg. Both figures include the weight of the end effector, so installing a heavier gripper reduces the available payload accordingly.

The most immediate hardware changes in 2.0 focus on the end effector. The new parallel gripper uses a more compact actuation mechanism, features replaceable fingers, and integrates a camera into the gripper housing to capture close-range images during grasping. This helps collect manipulation data involving severe occlusion and fine contact, while allowing researchers to modify the fingers for different object geometries without redesigning the entire gripper. The tradeoff is that different fingers, camera viewpoints, and end-effector weights can still alter the policy distribution, so datasets must fully document the hardware configuration.

The more important new component is the OpenArm Cell. By standardizing the background, lighting, camera positions, calibration method, and arm mounting points, it reduces the domain shift caused by laboratories building their own experimental setups. The public repositories also provide ROS 2, CAN control, MuJoCo, Isaac Lab, teleoperation, a data logging format, and a Python API. The hardware CAD is licensed under CERN-OHL-S-2.0, while the main software uses Apache 2.0. This approach—releasing the mechanism, simulation, data format, and test environment together—comes closer to reproducible robot-learning infrastructure than merely publishing a URDF.

However, a standardized setup can control only some variables and cannot yet demonstrate that different units, motor batches, camera exposure settings, or assembly tolerances will produce equivalent results. The project has not yet published a large-scale, multi-site benchmark using the Cell, calibration error ranges, or policy success rates. The planned motorless KER passive teleoperation device has also not yet been fully released. The hardware repository further notes that Google Drive, rather than GitHub releases, is the source of truth for the CAD and BOM, with GitHub serving only as a mirror. Research teams should next focus on version pinning, sensor calibration files, the data schema, and real-hardware testing protocols. Otherwise, systems bearing the same “OpenArm 2.0” name may still produce results that are not directly comparable.