Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

MABEL: An Open Anthropomorphic Mobile Bimanual Platform for Teleoperation and Learning

Qilong Cheng, Rooholla Khorrambakht, Mohamed Yassine Kabouri, Quang Nam Nguyen, Luca Macesanu, Haoran Zhou, Yipeng Wang, Ludovic Righetti

November 2026

Teleoperated household tasks: opening a fridge, retrieving from a high shelf and closing a dishwasher.

TL;DR An open-source 56-DOF mobile bimanual robot with dexterous hands, a pitching torso, a lift and an active neck, for a US$8.1k parts cost. Apple Vision Pro head and hand tracking drives the whole body with 137 ms video latency, and an ACT policy trained through the same controller succeeds in 8 of 10 grasp-and-place trials.

What they do

Commercial robots that combine mobility, floor-to-overhead reach, active gaze and dexterous hands cost US$69k-120k. MABEL packs all of these into a fully open design with a US$8,056 parts cost. CAD, firmware, ROS 2 stack, controller, retargeter, learning pipeline and a MuJoCo and Isaac Lab twin are all released. The authors built a second unit from the released files in about two weeks.

How it works

  • Hardware. Two back-drivable 7-DOF arms adapted from OpenArm carry 17-DOF ORCA hands (swappable for grippers). A quasi-direct-drive torso pitches from -60° to 15°, a desk lift adds 0.635 m of travel, a 3-DOF neck carries a stereo head camera, and the base has three swerve modules.
  • Teleoperation. Vision Pro head and hand tracking alone sets the whole body: leaning sets torso pitch, raising or lowering the head sets lift height, walking moves the base. Arm IK scales for reach, and finger retargeting preserves pinch apertures.
  • Shared control. Teleoperation, simulation and policies use one controller: collision-avoiding joint velocities, low-gain impedance with gravity and friction compensation, and base speed limits that tighten as the lift rises.

Results

  • Median palm repeatability is 1.65 mm and path error 3.7-6.7 mm RMS. Gravity feedforward cuts palm error from 194 mm to 45 mm under compliant gains.
  • Wheel odometry differs from visual-inertial tracking by 13.5-17.7 mm RMS.
  • Glass-to-glass latency is 137 ms on the robot’s hotspot, 150 ms on LAN and about 260 ms through a cloud relay. The camera accounts for 101-108 ms.
  • An ACT policy trained on 117 episodes (38 min) grasps an object and places it on a plate in 8 of 10 trials.

Limitations

The lift is slow, and OpenArm’s link proportions differ from a human’s. The policy demos are mostly upper-body, so they validate the pipeline but not mobile behavior. Long-horizon autonomy and onboard mapping are future work.

Videos

Local-mode teleoperation, where the robot's whole body mirrors a Vision Pro operator's head and hands.
Compliant whole-body control next to a person in a home, at 3.1x speed.

Figures

From Figure 1: MABEL, a 56-DOF open-source anthropomorphic mobile bimanual robot with a parts cost of US$8.1k.
From Figure 1: MABEL, a 56-DOF open-source anthropomorphic mobile bimanual robot with a parts cost of US$8.1k.
From Figure 3: Software stack. Teleoperation clients, simulation and learned policies share one command interface and...
From Figure 3: Software stack. Teleoperation clients, simulation and learned policies share one command interface and whole-body controller, which drives either the real robot or its digital twin.
From Figure 2: (a) Kinematic fingertip workspace. (b) Camera view of a fixed camera versus with neck and torso motion...
From Figure 2: (a) Kinematic fingertip workspace. (b) Camera view of a fixed camera versus with neck and torso motion. (c) Fingertip-height ranges against other platforms and a human reference, from joint limits only.
From Figure 8: Hardware repeatability and path following. (a) Ten return cycles, with covariance ellipsoids magnified...
From Figure 8: Hardware repeatability and path following. (a) Ten return cycles, with covariance ellipsoids magnified 30x. (b) Joint contributions to scatter. (c) Returns at true scale. (d) Measured traces against dashed references. (e) Tracking error over three trials.
Abstract

As robot learning moves beyond the tabletop, many tasks require mobility, adjustable reach, active perception, and bimanual dexterity. Commercial platforms combining these capabilities are costly, while affordable open-source alternatives cover only a subset and require software integration. We present MABEL, an open-source anthropomorphic mobile bimanual robot with 56 degrees of freedom (DOF) and a US$8.1k parts cost. It combines two 7-DOF arms with 17-DOF hands, a pitching torso, a lift, a 3-DOF neck, and a holonomic base. Back-drivable arms and torso support compliant interaction, with fingertip reach from floor level to 2.23 m. The head and torso roughly triple modeled camera coverage relative to a torso-fixed camera. Apple Vision Pro head and hand tracking enables whole-body teleoperation without handheld controllers, with 137 ms mean glass-to-glass stereo video latency over the robot's Wi-Fi hotspot. We demonstrate an end-to-end imitation learning pipeline, from teleoperated data collection to policy deployment, using a shared command interface and whole-body controller.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.