Organizations: University of Toronto, Toronto, ON, Canada · Center for Robotics and Embodied Intelligence (CREO), New York University, New York, NY, USA · LAAS-CNRS, Université de Toulouse, CNRS, Toulouse, France
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.
Figures & tables
Fig. 1: MABEL, a 56-DOF open-source anthropomorphic mobile bimanual robot with a parts cost of US$8.1k. Hardware designs, firmware, control software, simulation models, and the learning pipeline are available at robotmabel.github.io/website .
Fig. 2: Platform capabilities, cost, reach, and camera coverage. MABEL combines a holonomic mobile base, two 7-DOF arms, 17-DOF hands, a 3-DOF neck, a lift, and a pitching torso at roughly a tenth of the cost of comparable commercial platforms. (a) Kinematic fingertip workspace. (b) Modeled fixed-camera view and cumulative neck/torso visibility ( 60 ° vertical field of view (FOV)). (c) Fingertip-height ranges against other platforms and a human reference, from joint limits only.
Fig. 3: MABEL 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. HAL: hardware abstraction layer; QP: quadratic program.
[][] Parameter
Value
Parameter
Value
Functional DOF
56 (59 actuators)
Mass
61.8kg
Overall height
1.16m1.80m
Chassis
0.49×0.49\text{,}\mathrm{m}$$
Support polygon
0.34×0.35\text{,}\mathrm{m}$$
Model load limit
∼3.5kg
Parts subtotal
$8,056 (Pi 5)
Battery life
∼4h
TABLE I: MABEL specifications.
Fig. 4: Annotated cutaway of MABEL, showing actuators, sensors, compute, power, and electronics; colors group components by function.
Fig. 5: (a) Operator–robot pose correspondence, labeled frames, and magnified hand detail. (b) Joystick translation, lift, and yaw. (c) First- and third-person interfaces with numbered controls; the third-person background is dimmed. Arrows indicate head direction, torso pitch, lift, and base motion.
Fig. 6: Teleoperated household tasks on the physical robot, one continuous trial per row, time left to right. Arrows mark lift and door motion, the dashed arrow the gaze, the dotted path the hand, and teal arrows contact forces. Each label names the body part the step relies on: ∙ head, ∙ torso, ∙ lift, ∙ base, and ∙ arms and hands. Row (f) shows compliance: a person lifts, pushes, and pulls the robot by hand ( Sec. IV-D ).
Fig. 7: ACT learning pipeline. Head and wrist images It and measured joint states qt condition the prediction of a joint-reference chunk qt:t+K−1∗ , supervised during training by the demonstrated references qt:t+K−1∗ . The scene shows the pick-and-place task used for data collection and deployment.
Fig. 8: Hardware repeatability and path following illustration and results. (a) Ten cycles, with return samples and 95% covariance ellipsoids magnified 30× . (b) Joint contributions to scatter. (c) True-scale returns; 1mm rings, dashed median repeatability (RP). (d) Twelve measured traces; dashed references. (e) Tracking error: mean and range over three trials.
Fig. 9: Wheel odometry illustration: trajectories and signed x/y/yaw disagreements versus reference trajectory.
Fig. 10: Glass-to-glass video latency per connection method, from camera photon to operator pixel. Hotspot and LAN totals are measured; the relay total ( ≈ ) is partly derived from measured encode and transport times.
Teleoperation of mobile bimanual manipulators requires simultaneous control of high-dimensional systems, often necessitating expensive specialized equipment. We present an open-source teleoperation framework that enables intuitive whole body control using readily available commodity hardware. Our system combines smartphone-based head tracking for camera control, leader arms for bilateral manipulation, and foot pedals for hands-free base navigation. Using a standard smartphone with IMU and display, we eliminate the need for costly VR helmets while maintaining immersive visual feedback. The modular architecture integrates seamlessly with the XLeRobot framework, but can be easily adapted to other types of mobile manipulators. We validate our approach through user studies that demonstrate improved task performance and reduced cognitive load compared to keyboard-based control.
Danil Belov, Artem Erkhov, Yaroslav Savotin +3
Research Center for Digital Engineering and Innovation, Moscow, Russia
Mobile manipulation extends robot interaction beyond a fixed kinematic workspace by making the reachable region itself controllable. This flexibility introduces two central challenges: spatially grounded perception under continuous ego-motion and coordinated control of heterogeneous arm and base actions. Existing approaches strengthen geometry through explicit 3D representations or predictive world models, and often decouple mobility and manipulation into separate action streams. We argue that effective mobile manipulation requires not only decoupling, but also representations that support efficient cross-stream collaboration. We present MM-ABC, a foundation model built around Seeing, Coordinating, and Imagining Arm-Base Collaboration. MM-ABC combines sparse multi-level VLM features for spatial perception; a training-only future branch that uses world imagination and geometric intent as extra supervision, strengthening perception and manipulation-intent prediction and improving the overall learning signal; and MM-APT, which coordinates separate manipulation and mobility streams through masked joint attention and clean-action x-prediction. In controlled ablations, replacing clean-action prediction with velocity prediction lowers success on RoboCasa365 composite-seen tasks from 32.8% to 29.2%, and removing future supervision or multilevel conditioning causes larger drops. We pretrain MM-ABC on 5,000+ hours of heterogeneous robot data spanning 400K+ episodes, 12 datasets, and 17 embodiments. Experiments cover EBench, RoboCasa365, ManiSkill-HAB, LIBERO, LIBERO-Plus, and real-world mobile manipulation. MM-ABC achieves 44.71% success on EBench, 61.2% on RoboCasa365, 99.1% on LIBERO, 82.8% on LIBERO-Plus without perturbation training, and 83% mean success on five real-world tasks.
Qiwei Liang, Guangyu Chen, Shaolong Zhu +6
Xspark AI · The Hong Kong University of Science and Technology (Guangzhou) · Tsinghua University +1
The scarcity of large-scale, high-quality demonstration data remains a bottleneck in scaling imitation learning for robotic manipulation. We present COBALT, a teleoperation platform designed to democratize robot learning at scale both in simulation and in the real world. By leveraging vectorized environments, our scalable, load-balanced infrastructure supports concurrent teleoperation by multiple users on a single GPU, yielding a significant reduction in teleoperation cost. Operators can connect from nearly anywhere on Earth using commonly available devices, including single or dual smartphones, VR headsets, 3D mice, and keyboards. An inmemory data cache and efficient video streaming keep control and rendering synchronous, sustaining dozens of concurrent users at 20 Hz with sub-100 ms end-to-end latency for up to 8 concurrent users per GPU. We also demonstrate stable operation supporting 256 simulated clients across 8 GPUs, underscoring the system's ability to scale across hardware and within individual servers. We perform a comprehensive user study showing that phone-based teleoperation performs comparably to or better than specialized hardware, enabling faster, more ergonomic data collection. To ensure data quality, COBALT logs a suite of real-time metrics to automatically filter suboptimal demonstrations. We further demonstrate that a structured user training curriculum significantly improves data collection quality. Guided by insights from our user study, we crowdsource the collection of a large-scale, high-quality pilot dataset with 7500+ demonstrations (50+ hours) collected with smartphones across nine countries over five days. We validate the dataset's quality by training state-of-the-art imitation learning algorithms. Please visit https://cobalt-teleop.github.io/ for more details.
Ayush Agarwal, Ansh Gandhi, Jeremy A. Collins +6
Georgia Institute of Technology · University of California, Berkeley · New York University Abu Dhabi (NYUAD) +2