Source-linked AI summary
HaptiNet: Networked Haptic Robots Enable Physical Co-presence in Geographically-Unconstrained Rehabilitation
Chenyang Sun, Mingjie Dong, Haodong Deng, Yudong Liu, Yi-Feng Chen, Jun Lin, Changlong Huang, Jie Guo, Yantong Liu, Yang Liu, Yuzhou Lin, Jianjun Long, Zheng Xing, Sining Zhao, Xuemin Zhang, Zhiyong Wang, Zhenhong Li, Dongrui Wu, Honghai Liu, Jian S. Dai, Mingming Zhang
TL;DR
Cooperative rehabilitation is difficult to extend across distance because remote users lack force-mediated physical co-presence. HaptiNet connects force-feedback rehabilitation robots with imitation-learning-based delay compensation, and it maintained haptic interaction while improving task, engagement, synchrony, and clinical training outcomes. The authors caution that immediate training improvements should not be equated with long-term motor recovery.
Problem
Remote rehabilitation remains predominantly audiovisual, leaving geographically separated users without the haptic contact needed for physical co-presence and coordinated cooperative training.
Method
HaptiNet combines distributed low-inertia, long-stroke haptic terminals with a shared virtual environment and imitation-learning motion prediction for delayed remote states.
Results
HaptiNet preserved force consistency across user numbers and improved task performance, engagement, interpersonal synchrony, and patient training outcomes across healthy and neurological-impairment cohorts.
Takeaways & Limitations
HaptiNet supports geographically distributed cooperative rehabilitation with stable haptic interaction and measurable behavioral, physiological, and patient-training benefits.
Takeaways & Limitations
The results reflect immediate training performance and effort rather than long-term motor recovery; larger, longer, more rigorous randomized studies are needed.
Abstract
from arXiv · showhide
Cooperative rehabilitation enhances engagement, task performance, and social-motor interaction, yet it demands physical co-presence: users must transmit forces, coordinate movements, and infer intent through haptic contact. Telerehabilitation promises to expand access for patients constrained by distance, mobility, or clinical disparities, yet current techniques remain predominantly audiovisual while leaving users haptically and physically isolated. Here, we introduce HaptiNet, a networked haptic robotic system enabling physical co-presence for geographically distributed users via force-mediated interaction. Each robotic terminal features a low-inertia, long-stroke design with high force-feedback capacity, tailored for haptic rendering in upper-limb training. Building on these terminals, HaptiNet creates a distributed haptic network with an imitation-learning-based delay compensator, enabling users to physically perceive and coordinate with one another over distance. We validated HaptiNet in 284 healthy participants and 111 patients with neurological impairments across progressively realistic settings, including laboratory tests, cross-city deployments, and clinical applications. HaptiNet preserved task-level force rendering consistency across single-user and multi-user scenarios. Compared with solo and visual cooperative training, haptic cooperation improved task performance by 24% and 22%, respectively, while also boosting engagement and interpersonal motor synchrony. Across three intercity links totaling approximately 4,000 km, HaptiNet maintained stable haptic interaction among patients with neurological impairments, producing a 3.87-fold greater baseline-to-training score improvement and a 106% higher patient-applied effort over the solo condition.
Main Text: INTRODUCTION
HaptiNet addresses the geographic and haptic limitations of cooperative rehabilitation by connecting low-inertia, force-feedback robots with latency compensation. Across progressively realistic evaluations, it preserved task-level force consistency and produced behavioral, physiological, and clinical cooperative benefits.
- Neurological impairment is linked to social isolation, while conventional rehabilitation remains predominantly individualistic and limited in social interaction.
- Remote cooperative rehabilitation is constrained by the need to transmit forces and coordinate movements despite geographic separation and network latency.
- HaptiNet combines low-inertia, long-stroke, high-force-feedback terminals with a shared virtual environment and imitation-learning motion prediction.
- Across 284 healthy participants and 111 patients, HaptiNet preserved force transmission and improved task performance, engagement, and interpersonal motor synchrony.Compared with solo training, reported improvements were 24% in task performance, 49% subjective engagement, 27% objective engagement, and 16% motor synchrony.
- Across single-user, two-user, and three-user conditions, task-required force remained consistent under 4 N, 10 N, and 20 N resistance levels.Trial-averaged absolute force errors remained within 4%, 2%, and 3%, respectively.
HaptiNet-mediated haptic coordination outperforms solo and vision-only conditions
HaptiNet-mediated haptic cooperation outperformed both solo operation and visual-only coordination in task execution, while adding engagement and realism beyond multi-user participation alone.
- The experiment separated benefits of haptic interaction from those of multi-user participation and shared visual feedback alone.
- Haptic co-op produced higher game scores and lower tracking errors than both Solo and Visual co-op.Game-score comparisons were significant versus Solo (P = 0.001) and Visual co-op (P = 0.005); tracking-error comparisons were P < 0.001 for both.
- Physical effort was greater in both multi-user conditions than in Solo, indicating that participation increased motor involvement.
- 4.59 ± 0.52 was the engagement rating for Haptic co-op, versus 3.63 ± 0.57 for Solo and 3.71 ± 0.65 for Visual co-op.
- Haptic co-op increased haptic realism and cooperation realism relative to Visual co-op while maintaining haptic realism comparable to Solo.Haptic realism was 4.46 ± 0.49 versus 3.73 ± 0.81, and cooperation realism was 4.70 ± 0.44 versus 3.86 ± 0.91.
Haptic interaction enhances physiological engagement and interpersonal synchrony
HaptiNet-mediated haptic cooperation increased frontal physiological engagement and interpersonal upper-limb muscle synchrony relative to Solo performance.
- EEG measured physiological engagement while sEMG measured interpersonal action synchrony during Solo and Haptic cooperative task performance.
- Normalized engagement index increased during Co-op at F3, F4, and their bilateral average across the 3–13 s task interval.Values changed from 0.93 to 1.06 at F3, 0.94 to 1.06 at F4, and 0.94 to 1.06 for the bilateral average.
- Co-op increased interparticipant sEMG coherence across multiple upper-limb muscles compared with Solo.In the 1–3 Hz band, coherence was consistently higher in six of eight analyzed muscles.
- The combined EEG and sEMG results showed enhanced physiological engagement and interpersonal motor synchrony during haptic cooperation.
Robust haptic interaction under network delays
Network latency degraded cooperative task performance and force transmission, while HaptiNet’s imitation-learning motion predictor reduced mismatch and better preserved haptic interaction under delay.
- Task score decreased and resultant force increased as latency rose from 0 ms to 50 ms.The tested conditions were 0 ms, 10 ms, 30 ms, and 50 ms, corresponding to local, LAN, MAN, and WAN communication.
- HaptiNet compared no prediction, constant-velocity prediction, and imitation-learning prediction to compensate for delayed remote motion.
- At 50 ms, ILMP more closely followed the true remote trajectory and suppressed reversal-related overshoot compared with NMP and CVMP.
- 0.193, 0.051, and 0.009 were the latency-dependent force-ratio slopes for NMP, CVMP, and ILMP, respectively.ILMP kept resultant force closest to the no-latency baseline as latency increased.
- At 30 ms, ILMP outperformed NMP on task score, and at 50 ms it outperformed both NMP and CVMP.At 50 ms, ILMP exceeded CVMP with P = 0.011.
Long-distance deployment of HaptiNet across 4000 km
HaptiNet maintained remote haptic cooperation across three intercity links with distinct network delays, and ILMP outperformed Solo and CVMP in task performance and perceived cooperation quality.
- Three intercity links exhibited approximately 34.5, 40.1, and 43.8 ms delays for A–B, A–C, and B–C connections, respectively.
- Co-op 2 using ILMP achieved the highest task scores during both training and post-training testing.Co-op 2 significantly outperformed Solo and Co-op 1 in both phases (P < 0.001 for each comparison).
- ILMP preserved engagement and haptic realism better than CVMP during long-distance cooperation.Co-op 2 produced higher engagement than Solo and Co-op 1, while Co-op 1 had lower haptic realism than Solo and Co-op 2.
- ILMP preserved resultant-force rendering at a level not significantly different from Solo, whereas CVMP differed significantly from both conditions.
- The remote protocol included baseline solo testing, randomized Solo and cooperative conditions, post-training testing, questionnaires, and force analysis.
Clinical validation in patients with neurological impairments
Clinical validation involved 111 patients with neurological impairments in solo and local or remote HaptiNet cooperation settings. Cooperative training produced larger task improvements and greater applied effort, while the authors caution that these findings reflect immediate training effects rather than long-term recovery.
- 111 patients with neurological impairments participated in solo or HaptiNet-mediated cooperative rehabilitation, including local and remote sessions.Most participants were post-stroke patients; other conditions included brain tumors, moyamoya disease, and related disorders.
- Co-op score improvement was 12.81 versus 3.32 for Solo, while tracking MAE reduction was 0.88 cm versus 0.33 cm.Baseline scores and tracking MAE did not differ significantly between groups.
- Co-op produced larger improvements than Solo in both task score and tracking accuracy as training progressed.
- The system architecture combined distributed low-inertia, long-stroke terminals with imitation-learning-based motion prediction for delay-compensated cooperation.
- Validation progressed from fundamental haptic rendering and transmission tests to network-delay experiments and clinical rehabilitation scenarios.
- In the clinical stage, Co-op improved from baseline by 23% versus 6% for Solo and produced 106% greater patient-applied effort.The Co-op improvement was approximately 3.87-fold greater than Solo.
- The results mainly reflect immediate training improvements and should not be equated directly with long-term motor functional recovery.The authors call for larger cohorts, longer interventions, and more rigorous randomized controlled designs.
- The study supports HaptiNet as a platform for cooperative rehabilitation beyond co-located settings.
MATERIALS AND METHODS System Development
HaptiNet integrates lightweight force-capable robotic terminals, local haptic control, shared virtual-environment synchronization, and an ILMP-based delay compensator for networked rehabilitation.
- Each terminal conveys a user’s interaction forces and motion to other connected users through four integrated architectural components.
- Robotic device: The robotic device uses a jointly optimized transmission mechanism and motion range to combine low inertia, high force output, and lightweight construction.Timing-belt transmission was selected to reduce reflected inertia and support a lightweight structure.
- Control framework for local haptic rendering: The local haptic control framework supports no-load, single-user loaded, and multi-user cooperative task modes.
- Shared virtual environment synchronization module: The shared virtual environment synchronizes local replicas of the task object and remote partner proxies across terminals.
- ILMP-based delay compensator: The ILMP-based delay compensator predicts current remote-partner states from delayed motion information to reduce latency-induced motion asynchrony.
Shared virtual environment synchronization
HaptiNet maintains a complete local representation of the shared multi-user task, combining user–object virtual coupling with synchronization among distributed object replicas.
- Each terminal maintains local representations of all users’ virtual proxies and a dynamic replica of the shared task object.
- Virtual coupling converts relative proxy–object motion into interaction forces applied to each local object replica.The coupling uses spring and damping coefficients.
- The task-level interaction force is computed by summing the virtual-coupling forces generated by all users.
- Synchronization coupling reduces state deviations among object replicas caused by network delay, packet loss, prediction error, and numerical integration error.Its correction force is normalized by 1/(n−1) across remote replicas.
- The synchronization correction maintains consistency without being interpreted as a task-level user interaction force.Object motion is primarily determined by the combined virtual-coupling forces, while synchronization preserves shared-state consistency.
Model training for ILMP
ILMP is trained on recorded user–object motion to predict future velocity trajectories from recent histories, supporting latency compensation in HaptiNet.
- ILMP learns task-related motion patterns from recorded user–object interactions and estimates future velocity trajectories from recent motion histories.
- Single-user loaded movement tasks provide training samples linking user motion, virtual object motion, and target trajectory information.Motion data are resampled to the 100 Hz network-update rate; the local control loop runs at 1000 Hz.
- Two independent MLP predictors estimate remote-partner motion and shared-object motion.
- The user-motion predictor uses position and velocity histories, box-position history, and the target-position sequence as inputs.
- Each predictor minimizes mean squared error between predicted and measured future velocity sequences.The predictors separately represent user-velocity and box-velocity prediction.
Online motion prediction
During online interaction, ILMP predicts short-horizon remote motion, interpolates it to the robot-control rate, and reconstructs latency-compensated states for local rendering and synchronization.
- At each control cycle, ILMP uses the latest received remote position and velocity to predict a short-horizon future velocity sequence.The study outputs ten predicted velocity samples.
- The predicted velocity sequence is interpolated from the 100 Hz network-update rate to the 1000 Hz local robot-control rate.
- The latency-compensated remote position is reconstructed by integrating interpolated predicted velocity from the latest received position.
- The compensated state corresponding to measured network latency is used first, followed by subsequent predictions until new network information arrives.The reconstructed position updates the local replica of the remote user or shared object for synchronization and haptic rendering.
Participants and experimental protocol
The study included healthy participants and patients with neurological impairments under an ethically approved protocol with informed consent.
- The study was approved by the Ethics Committee of Southern University of Science and Technology under approval number 20230095.
- Participants included 284 healthy individuals and 111 patients with neurological impairments.
- All participants provided written informed consent before the experiments.
Statistical analysis
Statistical analyses were conducted using MATLAB R2025b, with detailed methods, results, a statistical report, and raw data provided in supplementary materials and accompanying CSV files.
- Statistical analyses used MATLAB R2025b.
- Detailed statistical methods and results are available in the Supplementary Materials.
- A more detailed statistical report and raw data are provided as accompanying CSV files.