// the find
enactic/openarm
A fully open-source humanoid arm for physical AI research and deployment in contact-rich environments.
OpenArm is an open-source 7DOF robot arm targeting physical AI research — teleoperation, imitation learning, and sim-to-real tasks. At $6,500 for a bimanual system it sits between toy hobby arms and $50k+ commercial platforms. The repo itself is mostly a Docusaurus documentation site; the actual hardware CAD, ROS2 code, and simulation files live in separate sub-repositories.
- Genuinely open hardware under CERN-OHL-S-2.0 — full CAD (Fusion 360, STEP, STL), PCB Gerbers, and BOM are published, not just URDF stubs.
- Simulation coverage is real: separate repos for MuJoCo spec files, Isaac Lab training tasks, and URDF/xacro, which means sim-to-real pipelines have something to work with from day one.
- The 'OpenArm Cell' standardized environment concept (fixed camera placement, lighting, background) is a practical idea for making imitation learning datasets reproducible across labs.
- Well-structured documentation with step-by-step assembly photos per joint segment and a motor firmware update guide — uncommon thoroughness for a hardware project at this stage.
- The main repo is essentially just a docs website; everything interesting is fragmented across 8+ separate repos with no monorepo tooling, making it hard to track cross-cutting issues or pin compatible versions together.
- No published benchmarks or real-world task success rates — the claims about contact-rich performance and backdrivability are asserted but not backed by numbers, which matters when evaluating whether $6,500 buys you something meaningfully different from alternatives.
- Relies on DM series actuators sourced primarily through Chinese manufacturers; supply chain risk and long lead times are not addressed, and the 'verified manufacturers' page is essentially a vendor list without quality evaluation criteria.
- The software stack (ROS2 + MoveIt2 + CAN library + Dora + Isaac Lab + MuJoCo) spans many independent repos with no clear integration test or CI that runs end-to-end, so version skew between packages is a likely pain point for anyone setting this up fresh.