Long-horizon mobile manipulation needs whole-body demonstrations, and collecting them is the bottleneck. Keeping the robot in the loop ties the dataset to robot uptime, while recording human motion without it instead requires a conversion step that decides the joint-space action the policy learns. SuperSuit is a single wearable that both teleoperates a wheeled bimanual robot and, with the robot absent, records demonstrations for it. Its arms are mechanically aligned to the robot's, so recorded joint motion needs no inverse kinematics and no retargeting; one body-worn tracker drives the torso and the planar base through a deadband that keeps posture micro-sway out of the base command. Both roles emit the same 22-D command over the torso, planar base, two 7-DoF arms, and two grippers, with arm targets expressed as forward-looking increments of the observed joint state. On three real-robot tasks this lets active demonstrations replace most robot-in-the-loop collection without changing policy quality. With a 10-episode robot-grounded seed held fixed, the remaining 100 episodes can be recorded without the robot, whereas active-only training falls well short. SuperSuit moves the limit on dataset size from the robot to the operator.
Figures & tables
Figure 1: The SuperSuit Framework. Our untethered wearable interface translates human embodiment to whole-body robot control via isomorphic arm manipulation and deadbanded base locomotion. Both teleoperation and active collection produce demonstrations for imitation learning.
Figure 2: System Architecture of SuperSuit. The wearable signals are separated into a mechanical-arm stream for isomorphic arm mapping and a tracker stream for torso and base control. The tracker pose is converted to torso targets and fixed-frame planar-base velocities, followed by a velocity-level deadband. The commands drive the robot during teleoperation and are logged in both collection modes for the offline annotation pipeline, where an LLM combines Paraformer transcriptions with kinematic events to propose subtask boundaries that a human annotator verifies and corrects.
Figure 3: Remote Teleoperation. The operator drives the robot in real time with SuperSuit in (a) Pick-and-Place, (b) Blocks Collection, and (c) Crate Stacking.
Figure 4: Active Demonstration. Robot-free collection under the same whole-body action schema.
Table 2: Policy success rate and effective throughput under different training data compositions.
Table 8
Architecture
Pick-Place
Blocks
Stacking
ACT [ 40 ]
70%
30%
20%
π0.5 [ 2 ]
85%
60%
40%
π0.5+
85%
65%
50%
Table 5: Ablation on policy backbone and subtask annotations.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Component
Specification
Role in SuperSuit experiments
Left/right arms
7 DoF per arm
Bimanual manipulation and object transport
Left/right grippers
1 DoF parallel gripper per arm
Binary or continuous grasping command
Active torso
Lift, yaw, and pitch
Aligns upper-body workspace with the task scene
Mobile base
Omnidirectional vx , vy , ωz command
Planar locomotion during long-horizon tasks
Global camera
Chest-mounted RGB camera, 640×512 at 30 fps
Global scene observation
Wrist cameras
Left/right fisheye RGB cameras, 640×512 at 30 fps
Local manipulation observation near each gripper
Appendix
Table 6: Robot hardware summary.
Figure 7: Two-finger gripper pair. The SuperSuit-side gripper used for data collection and the matching robot-side gripper, which share finger geometry and wrist-camera placement.
Subsystem
Implementation
Notes
Arm interface
Isomorphic left/right wearable arms
Robot-aligned joints
Joint modules
Dynamixel XL330-M288-T servos
Joint logging
Gripper interface
Robot-matched gripper handles
Matched appearance
Spatial tracker
HTC Vive Tracker on a back-fixed bracket, with HTC Base Station 2.0
Base/torso tracking
Audio input
Headset microphone
Narration capture
Wearable mass
3.2 kg
Full wearable system
Appendix
Table 7: Wearable hardware summary.
Parameter
Meaning
Value
Kx
Forward velocity scale
1.0
Ky
Lateral velocity scale
1.0
Kω
Yaw velocity scale
1.0
ϵx
Forward deadband threshold
0.03 m/s
ϵy
Lateral deadband threshold
0.03 m/s
ϵω
Yaw deadband threshold
0.05 rad/s
Appendix
Table 8: Base-command parameters. Values are conservative engineering defaults for indoor mobile manipulation and can be adjusted for a specific base.
Figure 8: Subtask annotation interface. The annotator presents the task-level subtask vocabulary, batch pre-labeling controls, synchronized playback, timeline markers, and multi-view robot observations. Human annotators use this interface to verify model-proposed subtask boundaries and correct the language label associated with each interval.
Parameter
ACT
π0.5
π0.5+
Training steps
50,000
10,000
10,000
Batch size
256
256
256
Optimizer
AdamW
AdamW
AdamW
Learning rate
1×10−4
5×10−5
5×10−5
Language input
None
Task instruction
Task instruction
Auxiliary language supervision
None
None
Subtask annotations
Appendix
Table 10: Policy training hyperparameters.
Task
Randomization
Success criterion
Time Limit
Pick-and-Place
Object and box poses
Target objects placed in box
50 s
Blocks Collection
Block pose/shape; receptacle pose
Required blocks collected
180 s
Crate Stacking
Crate and target poses
Crate stably stacked
100 s
Appendix
Table 11: Task-level evaluation protocol.
Figure 9: Model inference rollout for Task 1. The sequence shows the inferred Pick-and-Place behavior, where the robot approaches the tabletop scene, grasps the objects, and places them into the box.
Figure 10: Model inference rollout for Task 3. The sequence shows the inferred Crate Stacking behavior, where the robot approaches the source crate, lifts and transports it to the target stack, places it, and retracts after completion.
Condition
Pick-Place
Blocks
Stacking
Teleop 0 / Active 110
7 [18.1,56.7]
2 [2.8,30.1]
1 [0.9,23.6]
Teleop 10 / Active 100
17 [64.0,94.8]
12 [38.7,78.1]
8 [21.9,61.3]
Teleop 110 / Active 0
17 [64.0,94.8]
13 [43.3,81.9]
8 [21.9,61.3]
BRS-style
16 [58.4,91.9]
13 [43.3,81.9]
7 [18.1,56.7]
Appendix
Table 12: Best-checkpoint outcomes (successes/20) and 95% Wilson intervals.
Dexterous manipulation requires both large-scale task progression and precise contact-rich interaction, making it challenging to collect demonstrations that effectively support both regimes. We present SkillWeave, a heterogeneous demonstration framework for long-horizon dexterous manipulation that combines teleoperation for coarse reaching and transport with kinesthetic teaching for precise, contact-rich skills. To address the visual mismatch introduced by the demonstrator's presence during kinesthetic data collection, we propose an object-mask-conditioned diffusion policy that uses offline object segmentation for training supervision and a lightweight learned mask predictor at deployment, avoiding online segmentation and image inpainting. To mitigate distribution shift between independently trained sub-task policies, we introduce successor-aware terminal steering, which selects among actions sampled from the predecessor policy to guide the system toward states supported by the successor's demonstrated initial-state distribution. Across three real-world long-horizon tasks, SkillWeave achieves 27% average end-to-end success. Mask-conditioned kinesthetic policies improve dexterous sub-task success to an average of 65%, while successor-aware handoffs achieve an average composition efficiency of 87%. These results show that matching demonstration modality to interaction regime, explicitly addressing kinesthetic visual mismatch, and steering policy handoffs toward successor-supported states substantially improves long-horizon dexterous manipulation. Videos and code are available at skillweave-authors.github.io .
Ryosei Tamura, Xiaoxiang Dong, Uksang Yoo +4
Robotics Institute, Carnegie Mellon University, USA. · Department of Computer Science, Keio University, Japan. · College of Engineering, University of California, Berkeley, USA. +1
We present Whole-Body Mobile Manipulation Interface (HoMMI), a data collection and policy learning framework that learns whole-body mobile manipulation directly from robot-free human demonstrations. We augment UMI interfaces with egocentric sensing to capture the global context required for mobile manipulation, enabling portable, robot-free, and scalable data collection. However, naively incorporating egocentric sensing introduces a larger human-to-robot embodiment gap in both observation and action spaces, making policy transfer difficult. We explicitly bridge this gap with a cross-embodiment hand-eye policy design, including an embodiment agnostic visual representation; a relaxed head action representation; and a whole-body controller that realizes hand-eye trajectories through coordinated whole-body motion under robot-specific physical constraints. Together, these enable long-horizon mobile manipulation tasks requiring bimanual and whole-body coordination, navigation, and active perception. Results are best viewed on: https://hommi-robot.github.io
Mobile imitation learning on portable demonstration interfaces faces two coupled bottlenecks: locomotion-contaminated action labels and inference-induced execution latency on a continuously moving base. Recent wrist-mounted interfaces lower the cost of tabletop data collection, yet a single wrist view does not capture the global context required for base navigation. Adding a body-mounted camera entangles human walking with hand motion. Meanwhile, generative policies introduce hundreds of milliseconds of inference latency, during which the base advances past predicted waypoints, forcing backward corrections at action splices. This paper presents Mobile UMI, a hardware-free demonstration framework that addresses both gaps through three components. First, a dual-camera capture system records chest-centric global context and wrist-centric local interaction without any robot present. Second, a one-shot ChArUco-based spatial anchor unifies the chest and hand visual-inertial frames; the hand pose is then re-expressed relative to the chest to extract decoupled SE(3) manipulation and SE(2) base trajectories. Third, an asynchronous receding-horizon executor performs online state matching: each generated action chunk is realigned with the current physical pose so that expired waypoints are discarded before execution. The full system is evaluated on four long-horizon household tasks, achieving an average success rate of 83.8% over 100 trials per task. Controlled comparisons against ACT and Diffusion Policy show that the chest-relative label alone closes much of the gap; online state matching closes the remainder. These results indicate that, for mobile imitation learning under the tested conditions, explicit kinematic factorization combined with state-level latency alignment provides an effective solution without requiring architectural changes to the underlying policy class.