A wearable input/output interface learns personalized models of haptic perception

Wen Li, E. Magnus Adams, Zijie Xu, Jiawen Yao, Aybala Esmer, Xiaoyue Ni, Bozhi Tian, Pedro LopesarXiv 2026 · arXiv

TL;DR

Researchers have developed a wearable haptic interface that "closes the loop" by integrating a 36-channel electrotactile array with a capacitive touch sensor. By using a few human-in-the-loop calibration trials, the system builds personalized perceptual models that account for individual anatomical differences, allowing it to accurately render tactile and thermal sensations exactly where a user expects to feel them.

What Problem Does This Paper Address?

Most wearable haptic devices operate on an "actuator-centered" assumption: if you activate an actuator at coordinate $(x, y)$, the user will feel the sensation at $(x, y)$. In reality, human perception is far messier. Differences in skin thickness, impedance, and peripheral nerve branching mean that two people might perceive the same electrical stimulus in completely different locations—sometimes centimeters away from the actual electrode.

This problem is even more acute for thermal referral, a phenomenon where a user perceives heat at the site of a tactile stimulus rather than at the physical location of the heat source (like a Peltier element). Until now, mapping these personalized "perceptual maps" required exhaustive, time-consuming calibration that made practical applications nearly impossible.

Methods

The authors developed a multilayered conformal haptic sleeve (Figure 1) that combines:

  1. Output: A 6×6 electrotactile array and a Peltier element for thermal stimulation.
  2. Input: A capacitive touch layer that allows users to "draw" exactly where they feel a sensation directly onto the device.
Fig. 1. Overview of haptic I/O device architecture. (A) Schematic of the multilayer wearable interface, integrating an electrotactile stimulation array with a capacitive touch array and external

To make calibration efficient, they moved away from exhaustive mapping toward a perception-aware framework using two main models:

  • Tactile Model: A global-offset-corrected population-mean model. It uses just five calibration trials to calculate a spatial offset for a new user, applying this to a pre-existing "average" map of 23 other participants.
  • Thermal Model: A Bayesian framework that treats tactile percepts as a spatial cue. Since tactile and thermal percepts are often spatially coupled, the model uses tactile data to guide thermal predictions, significantly reducing the number of thermal trials needed (which are slow due to skin recovery times).

These models enable inverse rendering: given a target location in a virtual environment, the system selects the specific electrode pair most likely to produce that percept for that specific user.

Datasets

The study utilized two primary groups of participants:

  • Mapping Cohort: 24 participants (13 male, 11 female) used to establish the initial perceptual maps, characterize inter-participant variability, and train the predictive models.
  • Validation Cohort: 12 independent participants used to test if models learned from static points could generalize to "animations" (sequential stimulations).

The authors have made their hardware designs, firmware, and raw data available on GitHub.

Key Insights and Ideas

  • Shared Spatial Structure: While absolute percept locations vary wildly between people, the relative structure is preserved. If electrode A feels "above" electrode B for one person, it likely does for everyone, even if the whole map is shifted. This is why a population-mean model with a simple offset works so well.
  • Multimodal Coupling: The researchers found a significant electrode-pair-specific correspondence between tactile and thermal percepts. This "aha moment" allowed them to use fast tactile reports to predict slow thermal sensations.
  • Sparse Calibration is Sufficient: You don't need to test every electrode. The team demonstrated that just five trials are enough to outperform traditional "actuator-centered" rendering.

Key Results

  • Accuracy Gains: The population-mean approach significantly outperformed predictions based on electrode geometry ($P < 0.0001$). Adjusting for a global offset reduced tactile prediction error to 19.35 mm.
  • Thermal Prediction: Incorporating tactile guidance reduced thermal prediction error from 27.11 mm to 14.22 mm in participants with strong tactile-thermal alignment.
  • Inverse Rendering: The personalized models increased the "pair-selection agreement" (the likelihood of picking the "correct" electrode for a target) from 18.8% to 31.0%.
  • Real-world Applications: The authors demonstrated the system in four scenarios:
  • VR: Users "flick" virtual snowflakes off their arm to calibrate the system in-context.
  • Interpersonal Haptic Communication: Sending a "heart" shape drawn on one person's arm to another's.
  • Sensory Substitution: Using a thermal camera to help a user "feel" the location of a hot soldering iron.
  • Robot Teleoperation: Providing directional heat cues to a robot operator in a fire-rescue simulation.
Fig. 2. Characterization of the wearable haptic I/O system. (A) Circuit schematic of the wearable control hardware for electrotactile stimulation, high-voltage multiplexing, thermal actuation, and capacitive touch sensing. (B) Spatial accuracy of the capacitive touch layer measured with the device placed on a flat surface and conformally mounted on the forearm. Touch localization error was calculated as the Euclidean distance between the center of the targeted electrode and the measured touch location (36 electrodes × 3 repeats). Small points represent individual touch errors. Large circles represent run means. Whiskers extend to 1.5 × IQR. No significant (n.s.) difference was observed between the two conditions (two-sided paired Wilcoxon signed-rank test, N = 3 paired runs, P = 0.50). (C) Spatial map of electrode-to-array impedance. For each measurement, the target electrode was connected to the positive terminal, while the remaining 35 electrodes served as a common negative return. Impedance was measured at 4.096 kHz using a 3.20 µA excitation current. (D) Combined electrotactile-thermal stimulation profile. Thermal stimulation was applied for 5 s, while electrotactile stimulation was delivered for 7 s, extending 2 s beyond the thermal stimulus. Electrotactile stimulation was delivered with 3 mA amplitude, 100 Hz and 100 µs pulse width.

Limitations and Future Work

  • Anatomical Scope: The current study is limited to the ventral forearm. Perceptual maps for more complex areas (like the hand or back) may show different levels of variability.
  • Thermal Decay: Thermal sensations decay slowly, requiring a 30-second wait between calibration trials. While the tactile-guided model helps, thermal calibration is still the bottleneck.
  • Nerve Mapping: The authors suggest that future work could integrate high-resolution MRI of forearm nerves to see if specific branching patterns correlate with percept dispersion, potentially allowing for "zero-shot" calibration based on medical imaging.

Summary updated

Abstract

Wearable haptic interfaces deliver physical stimuli to the skin, yet perceived locations vary widely across users and diverge from actuator locations, particularly for emergent tactile-feedback techniques. We developed a wearable haptic input/output (I/O) device integrating a 36-channel electrotactile array with capacitive touch sensing, enabling stimulation and user-reported percept locations to be registered within the same body-referenced coordinate frame. In 24 participants, the device revealed electrotactile and referred-thermal perceptual maps with notable interparticipant variability. Despite this variability, both modalities retained electrode-pair-specific spatial structure across participants. Tactile and thermal percepts also showed spatial correspondence, whose strength varied across participants. These findings motivated our human-in-the-loop calibration framework, which adapts a forward-prediction strategy to individual users and generates personalized perceptual models. These models enabled inverse rendering by selecting stimulation patterns predicted to generate a target percept for a specific user. Using these models, we demonstrate personalized haptic rendering in haptic-guided robot teleoperation, visual-to-haptic sensory substitution, interpersonal haptic communication, and virtual reality (VR). Together, this platform provides a route toward adaptive haptic interfaces that measure, model, and render according to each user's perception.
BibTeX
@misc{li2026wearable,
  author = {Wen Li and E. Magnus Adams and Zijie Xu and Jiawen Yao and Aybala Esmer and Xiaoyue Ni and Bozhi Tian and Pedro Lopes},
  title = {{A wearable input/output interface learns personalized models of haptic perception}},
  year = {2026},
  eprint = {2611.09708},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09708},
}
Figure