SuperSuit: An Isomorphic Bimodal Interface for Scalable Mobile Manipulation

Abstract: High-quality, long-horizon demonstrations are essential for embodied AI, yet acquiring such data for tightly coupled wheeled mobile manipulators remains a fundamental bottleneck. Unlike fixed-base systems, mobile manipulators require continuous coordination between $SE(2)$ locomotion and precise manipulation, exposing limitations in existing teleoperation and wearable interfaces. We present \textbf{SuperSuit}, a bimodal data acquisition framework that supports both robot-in-the-loop teleoperation and active demonstration under a shared kinematic interface. Both modalities produce structurally identical joint-space trajectories, enabling direct data mixing without modifying downstream policies. For locomotion, SuperSuit maps natural human stepping to continuous planar base velocities, eliminating discrete command switches. For manipulation, it employs a strictly isomorphic wearable arm in both modes, while policy training is formulated in a shift-invariant delta-joint representation to mitigate calibration offsets and structural compliance without inverse kinematics. Real-world experiments on long-horizon mobile manipulation tasks show 2.6$\times$ higher demonstration throughput in active mode compared to a teleoperation baseline, comparable policy performance when substituting teleoperation data with active demonstrations at fixed dataset size, and monotonic performance improvement as active data volume increases. These results indicate that consistent kinematic representations across collection modalities enable scalable data acquisition for long-horizon mobile manipulation.

• The paper introduces SuperSuit, an isomorphic bimodal interface that combines robot-free wearable demonstrations with robot-in-the-loop teleoperation through a shared delta-joint action representation and continuous base control.
• SuperSuit increases active demonstration throughput by 2.5–2.9× over the BRS baseline and achieves 2.0–2.5× higher effective throughput when lower-cost active data replaces teleoperation episodes.
• The paper shows that delta-joint actions are critical for consistent learning, with success reaching 40% versus 5% for absolute-joint actions, while language-based subtask annotations improve long-horizon crate stacking by 10 percentage points.

Motivation and problem statement

Imitation learning policies for wheeled mobile manipulators require whole-body demonstrations that couple $SE(2)$ base locomotion with precise arm control, but existing acquisition interfaces impose severe constraints. Robot-in-the-loop teleoperation systems such as BRS and HOMIE decouple arm control from base motion via joysticks or pedals, fracture operator embodiment through mediated 2D camera views, and bind data throughput to robot availability. Robot-free wearable approaches such as HumanoidExo and EgoHumanoid rely on SLAM-based tracking whose accumulated drift misaligns precision manipulation trajectories, while UMI-style 6D end-effector tracking requires inverse kinematics with its attendant singularities and non-unique solutions. SuperSuit addresses these limitations with a bimodal wearable framework that supports both active human demonstration (robot-free) and robot-in-the-loop teleoperation under a single shared kinematic interface (2603.06280).

System architecture

The hardware consists of a lightweight 3D-printed isomorphic exoskeleton mirroring the target robot's arm kinematics, a head-mounted HTC Vive Tracker localized by lighthouse base stations at sub-millimeter accuracy, and a headset microphone capturing in-situ verbal narration. The software pipeline comprises a dual-stream control engine: an upper-body stream performing isomorphic joint mapping, and a tracker stream decomposing head pose into articulated torso configuration and planar base velocities.

Locomotion retargeting. After offline rigid calibration between the tracker frame and the torso pitch link, the torso lift, yaw, and pitch joints are read directly from vertical displacement and orientation components $(z, \psi, \theta)$. Planar velocities are obtained by discrete differentiation of the ground-projected pose, yielding continuous commands $(v_x, v_y, \omega_z)$ rather than discrete button-mediated switches. An adaptive velocity-level deadband suppresses involuntary postural micro-sway that would otherwise induce base jitter during contact-rich manipulation.

Manipulation formulation. Strict mechanical alignment of exoskeleton rotational axes with anatomical degrees of freedom establishes a bijective joint-space mapping that bypasses Cartesian transformations entirely. Rather than logging absolute joint targets—which are sensitive to calibration offsets, gear backlash, and structural compliance—the action space uses a shift-invariant forward-looking delta:

$$\mathbf{a}_t = \Delta \mathbf{q}_t = \mathbf{q}_{t+k} - \mathbf{q}_t$$

where $k$ is a look-ahead horizon aligned with the action chunk boundary. Constant offsets cancel under differencing, making trajectories structurally consistent across both collection modalities without requiring IK.

Language annotation. A Human-in-the-Loop pipeline transcribes narration audio with Paraformer, then uses Qwen3 as a kinematic reasoning engine to propose physical breakpoints (zero-velocity crossings, gripper toggles) and temporally align transcribed text to them; operators verify boundaries before subtasks are emitted as language-conditioned supervision.

Experimental results

Experiments use a custom 22-DoF wheeled bimanual humanoid (two 7-DoF arms with parallel grippers, 3-DoF torso, holonomic base) evaluated on three long-horizon tasks: Pick-and-Place, Blocks Collection (bimanual receptacle holding plus dynamic picking), and Crate Stacking (bimanual lift, transport, and stacking). Policies are trained on $\pi_{0.5}$ with batch size 256 for 10,000 steps; an augmented variant $\pi_{0.5}^{+}$ adds an auxiliary autoregressive language head weighted at 0.1.

Throughput

SuperSuit teleoperation outperforms an adapted BRS baseline by 14–17% across tasks, attributed primarily to continuous stepping-to-velocity locomotion eliminating micro-adjustment pauses. Active demonstration mode achieves 151.4, 51.9, and 32.5 episodes/hour versus BRS's 56.8, 20.7, and 11.4—a 2.5–2.9× throughput increase, largest on coordination-intensive tasks where remote sequential attention bottlenecks bimanual synchronization.

Data substitution and effective throughput

At a fixed dataset size of 110 episodes, replacing 100 teleoperation demonstrations with 100 active ones yields nearly identical success rates (85%/60%/40% vs. 85%/65%/40% across the three tasks), supporting the claim that cross-modal structural consistency permits direct data mixing. Because active episodes are cheaper to acquire, effective throughput—successful autonomous completions per hour of collection effort—improves systematically by 2.0–2.5× (e.g., Pick-and-Place rises from 41.9 to 88.8 successes/hour). The authors attribute this gain beyond acquisition cost to smoother, temporally more consistent active demonstrations producing shorter execution horizons and fewer recovery behaviors.

Scalability and ablations

Holding teleoperation data fixed at 10 episodes, Crate Stacking success improves monotonically with active data volume: 15% at 50 episodes, 40% at 100, 55% at 200, and 65% at 400, with no early saturation reported. Two ablations isolate design choices:

Ablation	Condition	Success rate
Action formulation	Absolute joint ($q$)	5%
	Delta joint ($\Delta q$)	40%
Subtask annotations ($\pi_{0.5}^{+}$)	Blocks / Stacking	65% / 50%
Baseline ($(z, \psi, \theta)$0)	Blocks / Stacking	60% / 40%

The absolute-to-delta ablation is the starkest result in the paper: success collapses from 40% to 5% when the delta representation is removed, indicating that static calibration offsets and embodiment inconsistencies dominate failure modes in long-horizon bimanual coordination. The subtask annotation ablation shows language-conditioned stage awareness yields +10% absolute on Crate Stacking while leaving short-horizon Pick-and-Place unchanged, suggesting temporal regularization benefits scale with task horizon.

Limitations and open questions

The paper concedes several constraints. The annotation pipeline remains human-in-the-loop: Qwen3-proposed breakpoints require operator verification, so fully automated segmentation is not demonstrated. The evaluation covers a single custom platform and three tasks with 20 trials per condition, leaving statistical robustness at low success rates (e.g., 15%) uncertain. The deadband thresholds and calibration are tuned per-platform, and the framework's extension to heterogeneous embodiments—identified by the authors as future work—is asserted rather than validated. Haptic feedback for contact-rich interaction is absent, so active-mode demonstrations lack force information during grasping. Whether the monotonic scaling trend continues past 400 episodes, and whether the delta-joint formulation transfers to platforms with different compliance characteristics, remain open empirical questions.

Conclusion

SuperSuit demonstrates that enforcing a shared, shift-invariant kinematic interface across robot-free and robot-in-the-loop collection modalities allows high-throughput active demonstrations to substitute for hardware-bound teleoperation without policy degradation, while delivering 2.5–2.9× raw throughput gains and monotonic scaling behavior. The delta-joint ablation indicates that representation choice, not merely data volume, is decisive for long-horizon mobile manipulation, and the integrated narration pipeline provides a practical route to language-conditioned subtask supervision for VLA training.