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

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.

• The paper introduces [PERSON_NAME], an anthropomorphic bimanual platform with 59 actuators (56 DOF) designed to support teleoperation, demonstration collection, simulation, and control learning; the platform costs US\)8,934.
• [PERSON_NAME] integrates two 7-DOF arms, 17-DOF tendon-driven hands, a pitching torso, vertical lift, 3-DOF neck, and a three-module holonomic swerve base; it improves mobility, reach, active perception, and compliance for mental-to-immersed human activity.
• The paper evaluates [PERSON_NAME] hardware through characterizations, reporting a median repeatability of 1.65 mm, a geometric path error of 3.7–6.7 mm, a stereo video latency of approximately 137 ms, and a compatibility test for learning policies (80% success rate in ACT demonstration experiment).
• follow_up_questions

Platform objectives and system contribution

The paper presents MABEL, an open-source anthropomorphic mobile bimanual robot intended to support teleoperation, demonstration collection, simulation, and learned whole-body control in settings that exceed tabletop manipulation. Its central design objective is to combine mobility, variable-height reach, active perception, compliant actuation, and dexterous bimanual manipulation without the cost and integration burden associated with commercial platforms. The resulting system has 56 functional degrees of freedom driven by 59 actuators and a reported parts cost of US$8,056, compared with approximately US$69,000–120,000 for commercial systems with partially comparable capability (2611.04083).

MABEL integrates two 7-DOF arms, two 17-DOF tendon-driven hands, a pitching torso, a vertical lift, a 3-DOF neck, and a three-module holonomic swerve base. The arms and torso are back-drivable and controlled through low-impedance joint-space impedance control, enabling physical compliance during contact-rich interactions. The lift and torso jointly provide a modeled fingertip workspace extending from floor level to 2.23 m. This combination is significant because the platform does not achieve its cost reduction by omitting the principal embodiment components required for mobile manipulation; instead, it uses relatively inexpensive actuators, off-the-shelf structural and electrical components, modular fabrication, and a reduced three-wheel base.

Figure 1: MABEL is a 56-DOF open-source anthropomorphic mobile bimanual robot with a reported parts cost of US$8.1k.

The authors position MABEL against open platforms that typically sacrifice one or more of these capabilities. Existing systems often use position-controlled servo arms, single-DOF grippers, fixed bases, rigidly mounted cameras, or limited arm workspaces. MABEL's contribution is therefore primarily architectural and integrative: it combines capabilities that are individually available in prior systems but rarely exposed through a single reproducible, open hardware and software stack.

Embodiment, workspace, and cost structure

The mechanical design is organized around modularity and relatively rapid replication. The platform has a mass of 61.8 kg, an adjustable overall height of 1.16–1.80 m, a 0.49 m square chassis, an approximately 3.5 kg modeled load limit per arm, and an estimated battery life of approximately four hours with the primary 24 V battery configuration. A second robot was assembled from the released design and bill of materials in approximately two weeks, although this assembly time does not eliminate the practical requirements of fabrication, calibration, sourcing, and integration.

The three-module swerve base provides holonomic motion while reducing drivetrain cost by roughly US$700 relative to a four-module configuration. The tradeoff is a smaller support polygon and consequently greater sensitivity to elevated centers of mass. MABEL addresses this through height-aware velocity limits based on a zero-moment-point approximation. Base acceleration commands are scaled according to lift height and estimated center-of-mass geometry, so the system becomes progressively more conservative as the lift raises the upper body. This is an engineering mitigation rather than a complete dynamic stability solution: the model neglects center-of-mass changes induced by upper-body posture and therefore cannot fully characterize aggressive whole-body motion.

The arms are adapted from OpenArm hardware, while the hands use 17-DOF ORCA mechanisms through modular wrist interfaces. The end effectors can consequently be replaced by grippers or other hands, which is important for experiments where dexterous manipulation is unnecessary or where the hand mechanics impose unacceptable data-collection or control complexity. The torso provides a pitch range of approximately $-60^\circ$ to $15^\circ$, and the lift adds 0.635 m of vertical travel.

Figure 2: MABEL combines holonomic mobility, two 7-DOF arms, 17-DOF hands, active neck motion, torso pitch, and a lift while providing floor-to-overhead fingertip reach.

The active neck and pitching torso are also treated as perception mechanisms rather than merely postural degrees of freedom. The authors report that their combined motion roughly triples modeled camera coverage relative to a torso-fixed camera. The head camera has a 63 mm stereo baseline intended to approximate human interpupillary distance, while synchronized wrist cameras provide close-range views for manipulation. This active-perception claim is based on modeled field-of-view coverage, not on a task-level comparison demonstrating improved success attributable specifically to neck and torso motion. Nevertheless, the design directly exposes camera viewpoint as a controllable variable in teleoperation and learning.

The electrical architecture uses custom Teensy-based boards, separate CAN buses for the base and upper body, Ethernet communication with the host, and ROS 2 for high-level control. The release includes CAD, fabrication files, electronics, firmware, ROS 2 software, simulation models, and learning infrastructure under CERN-OHL-S-2.0 and Apache-2.0 licenses. The completeness of this release is a substantive part of the contribution: the platform is presented not merely as a robot design, but as a reproducible experimental system.

Unified software and control architecture

MABEL's software stack is organized around a shared command interface and whole-body controller. Teleoperation clients, simulation, data recording, and learned policies issue commands through the same intermediate representation, after which the controller performs collision avoidance, kinematic projection, and low-level tracking for either the physical robot or its digital twin. This design reduces a common source of distribution shift in robot learning: demonstrations are collected through one execution pathway while autonomous policies use another.

Figure 3: Teleoperation, simulation, and learned policies share a command interface and whole-body controller that can drive either hardware or the digital twin.

The whole-body controller first converts desired end-effector, body, neck, hand, and base commands into joint references. Arm inverse kinematics uses position, orientation-related directional terms, and a rest-posture regularizer, subject to joint limits. The resulting arm references are projected through a velocity-level quadratic program that imposes joint-velocity limits and distance-based collision-avoidance constraints. The controller then tracks the safe references with impedance control.

For the arms and torso, the torque command includes gravity compensation, friction compensation, proportional stiffness, and damping. The use of low stiffness is enabled by back-drivable actuators and system identification intended to reduce model error. The resulting behavior permits a person to displace the arms and torso manually, after which the robot returns toward its commanded reference. The hands, neck, and lift use position control, so compliance is not uniform across the full body. This distinction matters for interpreting the system as compliant: the platform supports compliant upper-body interaction, but not a globally torque-controlled or uniformly force-controlled embodiment.

The base kinematics use the standard planar relation between body twist and each swerve module's velocity. Steering error is compensated by scaling drive speed, allowing the modules to reorient before full translational motion. The height-aware stability mechanism limits acceleration rather than directly optimizing whole-body dynamics. It is therefore appropriate for conservative teleoperation and reference tracking, but its behavior under fast manipulation, substantial payload shifts, or contact-induced disturbances remains uncharacterized.

Controller-free whole-body teleoperation

MABEL uses an Apple Vision Pro as the primary teleoperation device. The headset supplies head pose, wrist pose, and hand keypoints at 60 Hz, while stereo images from the robot's head camera are returned to the operator. The 63 mm camera baseline is matched to the nominal human interpupillary distance, and the images are displayed at approximately true angular scale. The interface consequently provides visual feedback and body tracking without handheld controllers.

The retargeting problem is intrinsically underdetermined: the Vision Pro tracks a sparse operator state, whereas MABEL has 56 functional DOF. The authors resolve this by defining a reduced operator representation and assigning distinct tracked quantities to body subsystems. Head displacement is decomposed into a bounded component interpreted as leaning and a residual interpreted as walking intent. Lean controls torso pitch, vertical head displacement controls lift height, and residual planar displacement controls the mobile base. Head orientation controls the neck relative to the estimated body-heading frame. Wrist position and orientation are mapped to arm and wrist configurations through scaled shoulder-relative coordinates, while finger joints are optimized to match fingertip positions and thumb-to-fingertip vectors.

This mapping is effective because MABEL is proportioned approximately to an adult operator, reducing the scale discrepancy between human motion and robot motion. It nevertheless depends on calibration, body-heading estimation, and anchor resets after recentering, clutch release, or tracking loss. The method does not reconstruct the operator's complete body state; it imposes a structured interpretation of sparse tracking signals. The use of on-screen joysticks for base and lift commands provides an alternative when the operator cannot physically walk or change vertical posture.

Figure 4: Vision Pro head and hand tracking is retargeted into arm, hand, neck, torso, lift, and base commands, with joystick controls available for constrained operating conditions.

The measured video latency is one of the paper's strongest systems results. Glass-to-glass latency over the robot's hotspot is $137 \pm 5$ ms, while LAN operation yields $150 \pm 0.2$ ms and direct peer-to-peer operation yields $161 \pm 16$ ms. A relay configuration is substantially worse, with an estimated total of approximately 260 ms. The camera pipeline dominates the local configurations, contributing 101–108 ms; encoding, decoding, and display contribute approximately 12 ms, while network transport contributes 25–31 ms locally. The result indicates that local immersive teleoperation is practical within the demonstrated configuration, but it also identifies the camera pipeline—not the command retargeting—as the principal latency bottleneck.

Figure 5: Glass-to-glass stereo video latency is approximately 137 ms over the robot hotspot and rises substantially when traffic passes through a relay.

The teleoperation demonstrations cover household interactions involving base motion, lift adjustment, active gaze, bimanual manipulation, door interaction, and compliant physical contact. These demonstrations establish that the command interface can coordinate multiple subsystems in continuous trials. They do not, however, provide a standardized quantitative comparison with alternative teleoperation interfaces in terms of task completion time, operator workload, or failure rate.

Figure 6: Continuous teleoperated household-task trials exercise MABEL's base, lift, neck, torso, arms, hands, and compliant interaction capabilities.

Simulation, data collection, and policy deployment

The paper provides MuJoCo and Isaac Lab digital twins containing the full 59-actuator configuration, simulated camera and LiDAR streams, and 35 static and dynamic scenes. The simulation bridge exposes the same command interface used by hardware, allowing teleoperation and policy execution to pass through the same control abstractions in both environments. This is a stronger integration claim than releasing only a URDF model: the simulation is intended to participate in the data-collection and policy-development loop.

The learning demonstration uses Action Chunking with Transformers (ACT). Demonstrations are collected with Vision Pro while recording head and wrist images, measured joint states, commanded joint references, base twists, lift commands, and timestamps. The policy receives the left head image, both wrist images, and a one-second history of joint positions. It predicts a four-second chunk of future joint references at 30 Hz, but executes only the first second before replanning. The predicted references are sent through the same whole-body controller used for teleoperation.

Figure 7: ACT predicts a chunk of future joint references from head and wrist images together with recent measured joint states, and the shared controller executes those references.

The end-to-end experiment focuses on grasping a small object and placing it on a plate with both dexterous hands installed. The dataset contains 80 demonstration takes segmented into 117 episodes, totaling 38.0 minutes and 68,364 frames. The split is performed by take, with 64 training, 8 validation, and 8 test takes. Training requires 8,000 steps, batch size 24, and approximately six hours on an Apple M1 Max.

Across ten autonomous trials with the object randomized within a 20 cm by 20 cm region and a 60 s time limit, the policy succeeds in eight trials. This 80% success rate validates the complete pipeline from immersive teleoperation through data curation, ACT training, reference prediction, whole-body control, and physical execution. The result should be interpreted narrowly. The task is short-horizon, the demonstrations are mostly upper-body, the base and lift remain near their initial configurations, and the evaluation uses only ten trials. Consequently, the experiment validates system integration more strongly than it establishes the general manipulation competence of the learned policy.

Hardware characterization

The hardware experiments quantify repeatability, trajectory tracking, odometry, and latency rather than relying solely on qualitative demonstrations. In the upper-body repeatability test, each palm visits four checkpoints over ten cycles. Median repeatability is 1.65 mm, with a best-case value of 0.46 mm. The torso contributes the largest identified joint-induced scatter, with 0.56 mm RMS per palm.

For mirrored 10 cm-radius spirals and 20 cm squares, geometric path error ranges from 3.7 to 6.7 mm RMS. Time-aligned tracking error is higher, at 15–19 mm RMS, because the palms lag their references; compensating for this lag reduces the error to 5–9 mm RMS. Gravity feedforward has a particularly large effect, reducing palm error under compliant gains from 194 mm to 45 mm. These measurements show that low-impedance control is compatible with useful Cartesian tracking, but also that feedforward modeling and temporal alignment are essential. The gap between geometric and time-aligned error indicates that controller dynamics, rather than only geometric calibration, materially constrain tracking performance.

Figure 8: Upper-body repeatability is millimetric, while path-following error is governed substantially by compliant-control lag and gravity compensation.

Wheel odometry is evaluated against visual-inertial reference measurements. On a repeated 5.72 m path traversed in 121 s, position and yaw disagreement are 13.5 mm RMS and $1.63^\circ$ RMS, respectively. On a 1.55 m turn-and-square trial lasting 35 s, the corresponding values are 17.7 mm and $1.47^\circ$. The authors argue that this centimeter-level local accuracy is sufficient for teleoperation sessions, where base references are expressed in the odometry frame. That conclusion is plausible for short-duration local tasks, but it does not establish robustness under long-horizon navigation, wheel slip, uneven floors, or accumulated drift.

Figure 9: Three-module swerve odometry tracks short local trajectories with centimeter-scale position disagreement and approximately 1.5-degree yaw disagreement.

Limitations and open questions

The paper explicitly identifies several limitations. The lift supplies substantial vertical reach but is slower and less anthropomorphic than a multi-joint lower body. Replacing it with a folding torso, waist, knee, and ankle structure could improve whole-body motion, but would change the cost, mechanical complexity, stability analysis, and control problem. The OpenArm link proportions also differ from human anatomy, limiting the fidelity of retargeting despite the generally anthropomorphic body layout.

The empirical learning evaluation is deliberately narrow. Base and lift motion are largely absent from the demonstrations, so the 80% success rate does not establish learned mobile manipulation. The policy predicts the full joint-reference representation, but the experiment does not isolate the contribution of the base, lift, neck, torso, or dexterous fingers. It also lacks comparisons against alternative policy architectures, demonstrations collected through different interfaces, or ablations of visual streams and state history.

Several system claims remain based on modeling or short experiments. Camera coverage is modeled rather than measured through task-conditioned perception metrics. The stability limiter neglects upper-body-induced center-of-mass motion. Odometry is evaluated over short local paths and does not characterize drift during extended operation. The latency measurement identifies the camera pipeline as dominant, but the paper does not report the corresponding end-to-end command-to-actuation latency or quantify the effect of latency on task performance. Finally, the release substantially improves reproducibility, but reproducibility of the reported cost and two-week assembly time depends on component availability, fabrication access, and calibration effort that are not fully quantified.

Specific open questions follow directly from these limitations: whether the shared controller preserves an advantage when demonstrations contain rapid base and lift motion; how dexterous-hand demonstrations compare with simpler grippers under identical tasks; whether active gaze improves policy success after controlling for additional observation bandwidth; and how whole-body policies perform when trained on the same platform across simulation, teleoperation, and autonomous data.

Conclusion

MABEL is presented as a complete open mobile-manipulation research platform rather than as an isolated mechanical design. Its principal contribution is the integration of a 56-DOF anthropomorphic embodiment, compliant arms and torso, dexterous hands, active perception, holonomic mobility, immersive whole-body teleoperation, full-body simulation, and a shared policy-execution interface at a reported parts cost of US$8,056. The strongest quantitative results are its 1.65 mm median repeatability, 3.7–6.7 mm geometric path error, approximately 137 ms local stereo-video latency, and 8/10 success rate on the demonstrated ACT task.

The paper's evidence supports the claim that MABEL is a technically coherent and reproducible platform for research in mobile bimanual manipulation and robot learning. It does not yet establish long-horizon mobile autonomy or broad task generalization. Those questions remain tied to the platform's unresolved limitations in whole-body policy data, dynamic stability, anthropomorphic lower-body motion, and extended navigation.