Physics-Informed Multimodal Framework for Estimating Human Arm Endpoint Impedance During Physical Human-Robot Interaction

Abstract

Humans can adapt their hand compliance dynamically according to task demands by modulating arm endpoint impedance through changes in arm configuration and muscle co-contraction. This work introduces a multimodal physics-informed machine learning framework for estimating human arm endpoint impedance during multi-degree-of-freedom interaction with a collaborative robot. In this pilot study, data were collected during static and dynamic interaction tasks with the robot. During each trial, muscle activity was recorded via surface electromyography (sEMG), joint kinematics were measured via motion capture (MoCap), interaction forces were recorded with a force sensor, and end-effector positions were obtained directly from the robot. The neural network models were trained to predict the impedance parameters identified from a perturbation-based experiment using these multimodal inputs. Two models were developed and compared: a Temporal Convolutional Network with Multi-Layer Perceptron head (TCN-MLP) as a baseline and a Temporal Convolutional Network Physics-Informed Neural Network (TCNPINN) that integrates physical consistency through a physics based loss term.
Results show that introducing the physics constraint improved the prediction accuracy of the inertia (M), damping (D), and stiffness (K) parameters compared to the purely data-driven model. The inclusion of dynamic movement trials preserved model stability and generalization. While the estimated parameters are not yet accurate enough for direct implementation, the limitations are analyzed and used to identify directions for achieving more consistent and robust results. Nonetheless, the findings indicate that the proposed physics-informed multimodal learning framework has strong potential for estimating human arm endpoint impedance during multi-joint, dynamic physical human-robot interaction.