Machine Learning-Enhanced Modular Ionic Skin for Broad-Spectrum Multimodal Discriminability in Bidirectional Human-Robot Interaction.

Abstract

Multimodal tactile perception systems that mimic the functionality of human skin are able to perceive complex external stimuli, facilitating advanced applications in human-machine interactions. However, current systems still struggle with limited sensing ranges and suboptimal decoupling strategies, restricting their effective multimodal sensing. To achieve broad-spectrum multimodal discriminability, a machine learning-enhanced modular ionic skin (MIS) is developed via a synergistic sensor-algorithm optimization strategy. From the sensing material perspective, process-controlled hard-segment modulation in the ionic gel enables the development of diverse ionic conductors with enhanced sensing properties: a minimum temperature coefficient of -4.00% °C⁻¹ (10-160 °C), a linear gauge factor of 2.95 (0-100%), and a maximum pressure sensitivity of 80.5 kPa⁻¹ (0-1.3 MPa). With respect to the decoupling algorithm, a data-driven decoupling model for the MIS is meticulously proposed and trained on a dedicated multi-stimuli dataset, achieving maximum decoupling ranges for temperature and pressure with prediction errors as low as 7.0%, while maintaining reliable strain detection despite temperature interference. The effectiveness and functionality of the system are demonstrated in a multimodal wearable hand kit for operator hand recognition and a robotic gripper kit for feedback, highlighting its potential in bidirectional human-robot interaction.