Inflatable Whole-Body Robotic Skin with Internal Time-of-Flight (ToF) Depth Sensing and Kinematics-Based Point-Cloud Prediction
Abstract
Whole-body physical human-robot interaction (pHRI) with humanoid robots is commonly addressed through collision avoidance and/or robotic skin, each with distinct limitations. Collision avoidance provides substantial geometric clearance but cannot handle or mitigate contact once it occurs. Robotic skin directly measures contact, yet thin-skin designs leave minimal safety margin; slow reactions or forces exceeding the skin's elastic limit may pose risks. We propose a robotic skin that combines a single-piece inflatable envelope with distributed time-of-flight (ToF) sensors. The inflatable skin creates a large standoff distance between the contact surface and the rigid frame, providing geometric margin comparable to collision avoidance. Sensing occurs within the controlled volume of the skin, reducing errors common in open-environment perception. Moreover, even small objects produce measurable deformation over a large area. ToF sensors with wide field-of-view achieve full coverage without blind spots, scaling to 116 sensors across a full-size humanoid robot at 20 Hz. To enable motion-agnostic contact localization, we train a neural network that predicts the nominal (contact-free) surface geometry from joint angles and detect external contact as deviations from this prediction. The system is validated in human–robot interaction experiments that include single-point contact, multi-touch, and distributed contact such as full-body hugging.
Key Contributions
Inflatable Robotic Skin
A single-piece commercial Baymax costume serves as the inflatable skin, physically separating the contact surface from the rigid frame through a large standoff distance. This provides geometric margin comparable to collision avoidance while enabling contact detection within the controlled volume of the skin.
ToF-Based Whole-Body Sensing
Custom multizone ToF sensor modules (VL53L5CX) with 45°-tilted orientation achieve 360° coverage around each body segment. A pipeline transforms raw distance arrays into a unified point cloud, covering the full humanoid body with 116 sensors across 16 modules at 20 Hz, without blind spots.
Learning-Based Contact Localization
A feedforward MLP predicts the nominal (contact-free) skin surface from joint angles. External contact is detected as deviations between the predicted and observed point clouds, and DBSCAN clustering rejects noise-induced false positives for robust contact localization during dynamic movement.
Interactive Demo
Adjust joint angles to see the AI-predicted nominal 3D whole-body point cloud in real time.
Initial load may take 1–2 minutes — the AI model (~30 MB) and 3D meshes are fetched from this server.
BibTeX
@article{Park2026BaymaxSkin,
title={Inflatable Whole-Body Robotic Skin with Internal Time-of-Flight (ToF) Depth Sensing and Kinematics-Based Point-Cloud Prediction},
author={Park, Juhee and Kim, Sunoo and Bang, Yunseong and Kang, Sangjin and Kim, Joohyung and Kim, Jung and Park, Kyungseo},
booktitle={Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2026},
url={https://williamsunookim.github.io/BaymaxSkin}
}