NYU releases MABEL: a mobile bimanual robot with 56 degrees of freedom for $8,056 in open-source parts

NYU publishes MABEL: a mobile bimanual robot with 56 degrees of freedom for $8,056 in open-source parts

🔎 Manipulation robotics just lost its favorite excuse

"You can't do bimanual manipulation research without a six-figure budget." That sentence, heard in dozens of European labs over the past few years, has just taken a blow to the knees. The Center for Robotics and Embodied Intelligence (CREO) at New York University has just published MABEL — Mobile Anthropomorphic Bimanual Embodiment for Learning — a fully open-source mobile bimanual robot whose complete bill of materials comes to $8,056 for the "Essential" configuration.

The figure deserves a closer look, because this is not some rickety prototype whose BOM forgets half the electronics. MABEL packs 56 functional degrees of freedom: two 7-DOF arms, two 17-DOF ORCA hands, an articulated torso, a 3-DOF neck, and a swerve-drive base. In the comparison published with the paper, the equivalent commercial platforms — Mobile ALOHA at $32,000, Rainbow RB-Y1 at $80,000, DexMate Vega at $98,000 — all stop at grippers or hands with a maximum of 6 DOF.

Why now? Because imitation learning robotics research is stagnating on a hardware bottleneck: the platforms capable of collecting high-quality demonstration data are too expensive, too closed, and too limited in dexterity. MABEL attacks all three problems at once, and the timing coincides with an open-source wave that is spilling over from software into hardware — as seen with research agents like ByteDance's DeerFlow or OpenSeeker-v2. But unlike those software projects, an open-source robot involves physical parts, mechanical tolerances, and lead times. That's where things get interesting.

The essentials

• MABEL is an open-source bimanual mobile robot from CREO (NYU), with 56 functional DOF: two 7-DOF arms, two 17-DOF ORCA hands, a torso with pitch and lift, a 3-DOF neck, and a swerve-drive base.
• The BOM for the "Essential" configuration is $8,056 (parts, prices as of July 31, 2026) — versus $32,000 for Mobile ALOHA, $80,000 for the Rainbow RB-Y1, and $98,000 for the DexMate Vega (comparison from the paper, October 2026 prices, not strictly equivalent in scope).
• Unlike all comparable platforms, MABEL retains 17-DOF hands where the competition ends up with grippers or hands with 6 DOF at most.
• The hardware is licensed under CERN-OHL-S-2.0, with all files published: STEP, fabrication, electronics, firmware, ROS 2 stack, and simulation.
• Reproducibility is proven: a second robot was assembled from the published files alone, in about 2 weeks.
• Whole-body teleoperation requires no handheld controller — it is done via body tracking, with an average glass-to-glass stereo video latency of 137 ms over the robot's Wi-Fi hotspot.
• The imitation learning pipeline is end-to-end: from teleoperated data collection to policy deployment, with a shared control interface.

Recommended Tools

What MABEL actually is — and what it changes

MABEL is a bimanual anthropomorphic mobile robot designed for one precise task: serving as a data collection platform for teleoperation and imitation learning. The paper (arXiv 2611.04083) describes a complete system — let's call it a robot — designed from the ground up so that a human can pilot it naturally, while it records everything needed to train a policy.

The specification that makes the difference is hand dexterity. The two ORCA hands total 17 DOF each. The paper's comparison is unambiguous: Mobile ALOHA, RB-Y1 and Vega all end up with a gripper or a hand with 6 DOF at most. Yet fine manipulation — turning a key, folding laundry, inserting a plug — comes down precisely to those degrees of freedom that commercial platforms lack. A lab that buys a $32,000 Mobile ALOHA collects push-pull data; a lab that assembles a MABEL collects anthropomorphic manipulation data.

Another architecture decision that matters: the arms are adapted from Enactic's OpenArm, with DAMIAO actuators and joint-space impedance controls that are backdrivable. Backdrivability means the human can physically guide the arm during teleoperation, which greatly simplifies collecting natural demonstrations. It's the kind of choice you only make in a project that thinks "data first."

The torso isn't decorative either: motorized pitch and lift, plus a 3-DOF neck that lets the robot orient its stereo camera the way a human turns their head. On a whole-body teleoperated platform, this isn't a comfort feature — it's additional demonstration information.

Mobility: a swerve-drive base, not a differential drive

The swerve-drive base is a rare choice at this price point, and it's significant. A classic differential constrains the robot's movements — it turns in place, it can't strafe, and its motion reference frame is less natural to teleoperate. Swerve-drive enables omnidirectional movement with wheel steering, which brings the robot's behavior closer to that of a human with a cart. For whole-body imitation learning, every gap between the teleoperator's morphology and the robot's is paid for in policy quality at deployment.

The $8,056 BOM: what's in it, and what isn't

Let's start with what the figure represents: the cost of parts for the "Essential" configuration, as of 31/07/2026. Not labor, not tooling, not an integrator's margin. It's an honest BOM insofar as the project has validated its own reproducibility — the strongest proof one can provide for a BOM: a second unit assembled from the published files, in about 2 weeks.

On parts price alone, MABEL costs 4 times less than Mobile ALOHA. But note: this comparison is deliberately unfavorable to competitors on one point — Mobile ALOHA at $32,000 is an assembled product, with support, warranty, and delivery, ready to use out of the box. The paper's comparison explicitly flags this — "outside an equivalent scope." A lab comparing the two has to ask itself what its own assembly labor will cost.

And that labor is not negligible. Two weeks to assemble a second robot from the files means two weeks of an engineer's or PhD student's time. At a realistic fully-loaded employee cost in a European lab, that represents several thousand euros in internal cost. The real cost of an assembled MABEL therefore most likely falls between $12,000 and $15,000 — still 2 to 6 times cheaper than the commercial alternatives, but not $8,056.

There's a second point the BOM doesn't capture: procurement logistics. The DAMIAO actuators come from a Chinese supplier whose lead times and sales terms for a European lab remain to be qualified. The ORCA 17-DOF hands are the component that most critically constrains the design — if Enactic changes its pricing or availability, the entire economics of the project shift.

The paper's comparison table, in plain terms

I'll note that the paper's table specifies that the competitors' prices are outside a strictly equivalent scope — in other words, we're comparing robots of different classes when it comes to dexterity. That's precisely the point: at comparable dexterity, there simply is no commercial offering.

The imitation learning pipeline: from teleoperation to policy, seamlessly

This is where MABEL sets itself apart from a mere open-source mechanical kit. The paper describes an end-to-end pipeline: from teleoperated data collection all the way to policy deployment, with a shared control interface and a whole-body controller.

In practice, teleoperation is done through body tracking — no gamepad, no exoskeletal master arm. The human moves, the robot follows. This is the most natural and information-dense mode of teleoperation: the demonstrations capture the use of the torso, the neck, and the legs of the base at the same time as the arms. A single whole-body controller orchestrates it all.

The figure that struck me most in the paper: the average glass-to-glass stereo video latency is 137 ms over the robot's Wi-Fi hotspot. Glass-to-glass means from the sensor (the camera lens) to the feedback display (the headset or the teleoperator's screen). At 137 ms, we're in the zone where the human can still perform fine gestures without increasingly pronounced divergence, but we're also close to the limit. It's a figure that deserves to be published and should be read as an honest indicator — lower latency would have required dedicated network infrastructure, which would have driven up the BOM.

This integration of teleoperation and learning into a single control interface is exactly what manipulation research has been calling for for years. Labs working on long-term agents like ByteDance's DeerFlow or on search agents like OpenSeeker-v2 illustrate the same trend on the software side: value is shifting toward integrated pipelines, not isolated components. MABEL does the same thing for embodied AI.

CERN-OHL-S-2.0: why the license choice is not a trivial one

The CERN-OHL-S-2.0 license is a copyleft license for hardware. In plain terms: you can use, modify, and redistribute, but if you distribute a modified version, you must publish it under the same license. This is not MIT, and it is not Apache 2.0. It is a political choice as much as a technical one.

For research, it is the right choice. A lab that improves the wrist, that changes the base's electronics, or that replaces the ORCA hands with a homegrown version is contractually incentivized to publish its variant. In an ecosystem where each lab has been reinventing its arms in silos for twenty years, this is the mechanism that can create a genuine common foundation.

The scope of what is published is what truly sets MABEL apart from façade "open-source" projects: all the design files (STEP), manufacturing files, electronics, firmware, the ROS 2 stack, and the simulation are published. STEP matters enough to call out, because it is the native format that mechanical engineers actually use. A repo that publishes STEP files rather than STLs is a repo meant to be modified, not just reprinted.

The simulation published with the robot is the second critical link in the imitation learning pipeline. A lab can start developing its policies in simulation, validate transferability, and only use the physical robot for data collection and final deployment. This is the working model that has become standard in modern robotics research — and it ties into a question VicOne recently highlighted in another way: in robotics, attack surfaces and testing surfaces play out in simulation just as much as in the physical world, as demonstrated by the discovery of 28 zero-days in under 48 hours with AI.

What this changes for European labs

The first effect is mechanical: the entry ticket to anthropomorphic bimanual manipulation with hand dexterity drops from an $80,000–98,000 budget to a bill of materials of $8,056. For a European lab, this means a PhD student can have their own manipulation robot, or a lab can afford a fleet of 3 to 5 MABELs for the price of a single commercial Vega. Now, data collection in imitation learning robotics is exactly the kind of task that parallelizes well: more robots collecting in parallel, more demonstrations, a deployable policy faster.

The second effect is structural: reproducibility of robotic experiments becomes real. When two labs work on the same paper with the same robot built from the same files, they're comparing policies on the same hardware. Today, reproducing a manipulation experiment between two labs equipped with different platforms is nearly impossible — you have to relearn the hardware, re-calibrate, adapt the interfaces. With MABEL, you're comparing software against software.

The third effect is more indirect: for European teams interested in AI applied to the physical world, it becomes possible to combine a local MABEL with open-source reasoning and planning models. A lab building a complete pipeline around a MABEL can simultaneously run its agents on open-source infrastructure — this is precisely the topic of open-source AI agents with Ollama running locally, and it applies just as much to conversational agents as to embodied agents that orchestrate a robot.

One real constraint remains, and it must be named: labor cost. Two weeks of assembly for a second unit is proof of reproducibility, not of productivity. A lab assembling its first MABEL should count on additional weeks for setup, calibration, and getting familiar with the stack. It's an investment in internal expertise, not just a parts purchase.

❌ Common Mistakes

Mistake 1: reading $8,056 as the price of the assembled robot

The "Essential" BOM as of July 31, 2026 includes neither labor, nor tools, nor setup time, nor component price fluctuations. Reproducibility was validated by a second robot assembled in ~2 weeks — starting from files, by a team that already knew the project. A lab starting from scratch should budget more.
The solution: budget a realistic cost of $12,000 to $15,000 including a realistic labor charge, and check the availability of DAMIAO actuators and ORCA hands before committing.

Mistake 2: comparing MABEL and Mobile ALOHA as if they did the same thing

The paper's comparison states it explicitly: the competing platforms fall outside a strictly equivalent scope. Mobile ALOHA has no 17-DOF hands, and a lab doing only pick-and-place has no need for that level of dexterity.
The solution: choose based on your research scope. For anthropomorphic fine manipulation, MABEL has no commercial equivalent. For coarse bimanual manipulation, a self-assembled Mobile ALOHA remains faster to deploy.

Mistake 3: neglecting the network constraint of teleoperation

137 ms of average glass-to-glass latency is the published figure — on the robot's Wi-Fi hotspot. A lab environment with saturated Wi-Fi or a complex network topology can significantly degrade that figure.
The solution: dedicate a network to the robot during data collection sessions, and measure your own glass-to-glass latency before blaming the imitation learning pipeline for noisy demonstrations.

Mistake 4: believing the CERN-OHL-S-2.0 license is "free like Apache"

CERN-OHL-S-2.0 is copyleft: a modified variant must be published under the same license. A lab that improves the robot for an industrial project must share back its modifications.
The solution: if the project leads to a product, plan the publication strategy or the negotiation of an alternative license with NYU from the very start.

Mistake 5: forgetting that the imitation learning pipeline is a chain, not a module

A lab that buys a MABEL and collects demonstrations with degraded latency, poorly configured camera coverage, or an inconsistent recording policy will get poor policies even with a perfect robot.
The solution: treat the entire chain — teleoperation, capture, dataset normalization, simulation, deployment — as a single system to instrument, the way VicOne does in robotic cybersecurity with its 28 zero-days discovered in less than 48 hours using AI.

❓ Frequently Asked Questions

Does the $8,056 price include labor?

No. That's the cost of parts for the "Essential" configuration, as of July 31, 2026. A second robot has been assembled from the published files alone in about 2 weeks, but that assumes a team already familiar with the project. A lab starting from scratch must factor assembly labor, bring-up, and calibration into its real budget.

What's the difference between MABEL and Mobile ALOHA?

Hand dexterity and the license. Mobile ALOHA ends up with 0-DOF grippers, whereas MABEL ships with two 17-DOF ORCA hands, plus 7-DOF arms with impedance control. Mobile ALOHA is also a fully assembled product at $32,000 (October 2026), whereas MABEL is an open-source BOM at $8,056 under CERN-OHL-S-2.0 that you assemble yourself.

Can you buy MABEL fully assembled?

No, not to my knowledge as of the publication date. All the hardware is published under CERN-OHL-S-2.0: STEP files, fabrication, electronics, firmware, ROS 2, and simulation. There's no indication that a commercial integrator currently offers an assembled version, and the copyleft license would require such an integrator to share their modifications back.

Is MABEL suitable for a lab outside the United States?

The project has proven its reproducibility by assembling a second unit from the files, which is the best indicator of geographic portability. The point of caution for a European lab is the supply chain logistics for the DAMIAO actuators and ORCA hands, whose lead times and sales terms outside Northeast Asia remain to be assessed.

What latency is needed for teleoperation with imitation learning?

The paper reports an average of 137 ms glass-to-glass over the robot's Wi-Fi hotspot, which is functional for teleoperation with stereo video feedback. That's close to the upper limit of comfort for fine gestures. A lab targeting highly precise tasks should aim for lower latency via a dedicated network.

Can the robot be used in simulation first?

Yes, the simulation is published along with the rest of the project, just like the ROS 2 stack, firmware, and STEP files. This lets you develop and test policies in simulation before consuming physical robot time — and use the actual MABEL only for demonstration collection and final deployment.

✅ Conclusion

With MABEL, NYU didn't publish just another robot — it published an open standard that brings the entry ticket for anthropomorphic bimanual manipulation down from six figures to just over $8,000 in parts, with the entire imitation learning chain documented end to end. The reproducibility test has already been passed: a second unit was assembled in two weeks from the published files alone. To the European labs that were waiting for affordable hardware to get into fine manipulation: the repo is there, the STEP files are there, and the rest is just a matter of ordering parts and two weeks in the workshop.