OpenArm: A $6,500 Open-Source Humanoid Arm for Physical AI Research

Most humanoid robotics hardware announced in 2026 has come from well-funded companies keeping their designs proprietary. Enactic's OpenArm is the opposite bet: a fully open-source, human-scale 7DOF robot arm — CAD files, control software, simulation environments, and dataset tooling all released under permissive licenses — with a complete bimanual system priced at $6,500.

That price point and the breadth of the software stack are what make OpenArm worth a closer look, especially for anyone doing physical AI research who doesn't have access to a proprietary humanoid platform or a six-figure robotics budget.

TL;DR: what OpenArm actually is

Why "contact-rich" and "human-scale" are the real pitch

Most industrial robot arms are built to be rigid and precise, which makes them bad at the thing humanoid research actually needs: safely touching, gripping, and manipulating objects and environments not perfectly known in advance. OpenArm is explicitly designed around high backdrivability and compliance — meaning the arm can be pushed or perturbed without fighting back destructively, a property that matters both for human safety around the arm and for tasks like inserting a plug, wiping a surface, or handing an object to a person, where some give in the joints is a feature, not a bug.

Human-scale proportions matter for a separate reason: most imitation-learning datasets for robot manipulation are collected via teleoperation, where a human operator controls the arm directly or through a matching controller rig. An arm built to roughly human dimensions makes teleoperation more intuitive and makes data collected on it more directly transferable to eventual humanoid deployment, compared to an arm with non-human link lengths or joint ranges.

The OpenArm Cell: a reproducibility play

One detail in Enactic's announcement stands out beyond the hardware spec sheet: "OpenArm Cell," a standardized environment with a unified background, consistent lighting, and fixed camera placement. The stated goal is that research performed using OpenArm can be reproduced in the same conditions anywhere in the world, which directly addresses one of the most persistent, underdiscussed problems in robot-learning research — results reported on one lab's custom rig rarely transfer cleanly to another lab's setup, because camera angles, lighting, and background clutter all affect what a vision-based policy actually learns. A shared reference cell design, if adopted broadly, would let OpenArm results function more like a genuine benchmark than a one-off demo.

The full stack: nine repositories, one pipeline

OpenArm isn't just a CAD file dump. Enactic has split the project into nine repositories that together cover the entire pipeline from physical hardware to trained policy:

• openarm_hardware — complete CAD data: STL files, STEP files, and Fusion 360 assemblies, under CERN-OHL-S-2.0
• openarm_description — URDF/xacro robot description files for simulation
• openarm_can — a CAN control library for low-level motor communication
• openarm_ros2 — ROS2 integration packages and nodes
• openarm_teleop — teleoperation packages supporting both unilateral and bilateral control
• openarm_isaac_lab — Isaac Lab simulation environment and training tasks
• openarm_mujoco — MuJoCo specification files and assets
• openarm_dataset — dataset format, recording tools, and a Python API
• dora-openarm — Dora dataflow nodes for data collection, inference, and teleoperation

For a researcher or independent builder, that's the difference between buying a robot arm and buying a research platform. The CAN library and ROS2 packages handle the low-level control problem that usually eats weeks of setup time on a new arm; the Isaac Lab and MuJoCo environments mean you can start training imitation-learning policies in simulation before your hardware even arrives; and the dataset format with a Python API addresses the unglamorous but critical problem of recording manipulation demonstrations in a format that's actually reusable across projects.

How it compares to YAM and other open arms

OpenArm enters a small but growing category of open-source research arms priced well below industrial robotics hardware. YAM, another open-source arm platform, has become a common reference rig in recent robot-learning benchmark reports, including disputed claims about GPT-6 Astra's manipulation performance. Compared to lower-cost single-purpose projects like the $399 Microduck open-source RL robot, OpenArm sits at a different tier — it's aimed at serious contact-rich manipulation research with a full bimanual system, not a minimal learning toy.

What differentiates OpenArm within this category is less the arm itself and more the completeness of the surrounding stack and the reproducibility framing of OpenArm Cell. A lab evaluating which open arm to standardize on should weigh: hardware cost and build complexity, existing community size (OpenArm's 3,100+ stars and 26 contributors indicate real, active use), simulation environment support, and whether a standardized test environment matters for the specific research question being asked.

This is also a broader pattern worth naming: 2026 has seen a real split between closed, well-funded humanoid platforms from companies like Figure and Unitree, and a parallel open-hardware track aimed at making physical AI research reproducible outside a handful of well-capitalized labs. Projects like Xiaomi-Robotics-1's 100,000-hour UMI pre-training release opened up data; OpenArm's contribution is opening up the hardware and the full control stack needed to collect more of that kind of data yourself, rather than depending on a dataset someone else already gathered on hardware you can't buy or reproduce.

Payload, backdrivability, and what "practical" means here

Enactic is specific about positioning OpenArm as practical rather than purely a research toy: the arm is meant to handle real payloads, not just move empty end-effectors through demonstration trajectories. That distinction matters because a lot of low-cost open-source arms trade off payload capacity to hit an aggressive price point, which limits what tasks they can actually be used to study — you can't research contact-rich manipulation of real objects, tools, or containers if the arm can't lift or grip them with enough force margin to do the job reliably. Combined with the compliance and backdrivability properties mentioned earlier, the goal is an arm that can both handle a workload and fail safely when it makes contact with something unexpected, which is the actual operating condition most manipulation research is trying to study in the first place.

The licensing split is worth understanding before you build

Enactic uses two different open licenses for a reason worth internalizing if you're planning to build on OpenArm commercially. CERN-OHL-S-2.0, applied to the hardware repository, is a "strongly reciprocal" open-hardware license — modifications and derivative hardware designs generally must also be released under the same license, similar in spirit to copyleft software licenses like the GPL. The software repositories, by contrast, use Apache-2.0, a permissive license that allows proprietary derivative works without the same reciprocal obligation. In practice, this means you can build proprietary software or research tooling on top of OpenArm's software stack freely, but if you modify and redistribute the physical arm design itself, those hardware changes generally need to stay open too.

What to watch next

Enactic says it's actively seeking research partners and company collaborators, and the project already has four tagged hardware releases, with the most recent being "OpenArm 01: Release No. 2." The real test for OpenArm, as with any open-hardware robotics project, is whether the manufacturer network listed on its site can reliably deliver assembled units at the advertised $6,500 price as demand grows, and whether enough independent labs adopt the OpenArm Cell standard to make cross-lab benchmark comparisons genuinely meaningful rather than aspirational. If both of those hold, OpenArm could become a real default reference platform for contact-rich manipulation research, the way certain open datasets and simulators have become defaults in other corners of machine learning.

There's also a question of long-term maintenance that any open-hardware project eventually has to answer: CAD files and control libraries need updates as motor suppliers change parts, firmware needs security and bug fixes, and simulation environments need to keep pace with new Isaac Lab and MuJoCo releases. A project with 26 contributors and active dependency-update commits is a healthier starting signal than a one-time hardware drop with no software follow-through, but sustaining that pace over multiple years — through component shortages, contributor turnover, and the inevitable slowdown after initial launch hype fades — is the harder, longer test that determines whether an open-hardware robotics project becomes genuinely durable infrastructure or a well-documented but eventually abandoned GitHub repository.

Related reading

• Reward AI's OM-1: a robot foundation model trained without robot data
• What is egocentric data collection for robots?
• GPT-6 Astra robot arm claims on YAM hardware, fact-checked
• HIW-500: BitRobot's open humanoid dataset for LeRobot
• Microduck: Hugging Face's $399 open-source RL robot
• DYNA-2: a world-action model and robotics scaling law
• NVIDIA Cosmos 3: an open physical AI world model
• Official source: OpenArm on GitHub, openarm.dev, docs.openarm.dev

Pricing, star counts, and repository details in this post reflect OpenArm's public GitHub listing as of September 15, 2026, and may change as the project develops.