Automatic Cognitive Fatigue Assessment Using Physiological Sensors during Human-Robot Interaction for Activities of Daily Living

Abstract

As assistive robots are increasingly integrated into home environments to support Activities of Daily Living (ADL), there is a critical need for systems that can assess and respond to users' Cognitive Fatigue (CF) in real-time. CF can impair performance and compromise safety, particularly during sustained Human-Robot Interaction (HRI). Despite its importance, existing approaches often lack multimodal sensing, continuous estimation, or adaptive interaction mechanisms. This study introduces a novel framework and dataset for automatic CF assessment using minimally intrusive physiological signals-Electrocardiography (ECG) and Electrodermal Activity (EDA),collected during collaborative ADL tasks. Data was gathered from 16 participants across up to three visits, incorporating N-back cognitive tasks to induce CF, and using the Visual Analogue Scale for Fatigue (VAS-F) as ground truth. Both epoch-based (45-second window with 5.625-second slide) and event-based machine learning pipelines were developed using Random Forest Regressor (RFR), Support Vector Regression (SVR), and Gradient Boosting Machine (GBM). The results show that RFR achieved the highest predictive accuracy across configurations, particularly with ECG alone (R-Squared (R 2) = 0.89, Root Mean Square Error (RMSE) = 4.13) and combined ECG+EDA (R 2 = 0.89, RMSE = 4.85) in epoch-based analysis. These findings demonstrate that continuous CF estimation is feasible using wearable sensors and regression modeling. This offline modeling forms the foundation for future real-time adaptation. The proposed framework lays the groundwork for integration into assistive robotic systems capable of dynamically adjusting interaction strategies in response to real-time CF levels-promoting personalized, safe, and inclusive HRI. The dataset is provided as supplementary materials to support reproducibility and future research.