MABEL
A whole-body robot you can actually build.
Capable mobile bimanual robots are proprietary; open ones are fixed-base with parallel-jaw grippers. MABEL is neither: 56 functional DOF — holonomic base, 0.635 m lift, actuated torso, active head, two 7-DOF arms, two 17-DOF hands — built for $9,670 with no machine shop, MIT-licensed end to end. One consumer headset is visualizer, controller, and demonstration recorder at once.
v1.0 · as built · no renders
Cheap, dexterous, whole-body
$9,670 as built. Two 17-DOF hands, two 7-DOF arms, a 3-DOF active head, and a holonomic swerve base — a 0–2.23 m fingertip band on a 0.49 m footprint. The only open platform that is holonomic, hand-dexterous, and neck-articulated under $10k.
Dexterity is normally where an open platform gives up. Hands get replaced by a parallel gripper, the neck gets bolted solid, the base gets two wheels and a caster. MABEL keeps all three and still lands under ten thousand dollars.
The swerve base is holonomic: it translates in any direction while the body keeps facing where the cameras are pointed, so the head never has to fight the wheels. The 3-DOF neck gives active perception — the robot looks where the task is instead of carrying a fixed camera and hoping.
Nothing needs a machine shop. Aluminium extrusion, laser-cut sheet metal, and printed parts, with the whole CAD tree MIT-licensed.
Modular by construction
Every subsystem is its own project on its own bus — swap an arm, a hand, the base, or the head without touching the rest. Extrusion + sheet metal; no machine shop; MIT-licensed CAD, firmware, and software.
Every subsystem is its own project on its own bus, behind its own connector. An arm is a unit you unbolt and carry away; the hand is a separate unit again; the head, the neck, the lift, and the base are four more.
That is what makes the platform a research base rather than a demo: swap in your own hand and the whole-body controller keeps working, because the controller talks to the kinematic model, not to a specific gripper.
Build one subsystem at a time. The bill of materials is split the same way, so you can price an arm, a hand, or the base on its own.
Simulation-complete
A MuJoCo twin runs the production control code unmodified at 500 Hz across 35 released
environments. Same URDF, same ROS 2 interface — use_sim:=true is the only difference.
Most robot simulators are a second implementation — a different controller, a different URDF, a different set of bugs. Here the twin runs the production control code, unmodified.
One MuJoCo MJCF is the single source of truth. The ROS URDF, the web viewer's GLB, the iOS and Vision Pro rigs, and the controller's tip-over constants are all generated from it, so the simulated robot cannot silently drift away from the real one.
use_sim:=true is the entire difference between driving the twin and
driving the hardware — same wire protocol, same safety gate, same ROS 2 interface.
One integrated system
Teleop (Vision Pro, iPhone, browser), electronics and firmware, data collection, and training are one stack: every command passes one wire protocol and one whole-body QP — and every session is a training demonstration.
The parts that usually live in four separate repos are one stack here: teleoperation, electronics and firmware, data collection, and policy training.
Every command source — Vision Pro, iPhone, the browser studio, a gamepad, or Nav2 — is a client of the same wire protocol, and every one of them passes the same whole-body quadratic program before a single joint moves. There is exactly one gate, so a safety fix lands everywhere at once.
And because teleoperation and data collection share that path, every session you fly is already a training demonstration — no separate recording rig.
