Mitigating Over-Assistance in Teleoperated Mobile Robots via Human-Centered Shared Autonomy: Leveraging Suboptimal Rationality Insights

Abstract

In this letter, we introduce a human-centered shared autonomy approach to address over-assistance in remote robot operation, aimed at reducing control conflicts and enhancing user experience. We model the human-robot team as a partially observable Markov decision process (POMDP) that incorporates uncertainties in intended goals and human rationality. By employing the Boltzmann noise-rationality model for predicting operator behavior and the regret theory-based mechanism for detecting model misalignment, we dynamically adjust assistance strategies to accommodate the operator's suboptimal rationality. Our experiments in three scenarios validate the proposed method, demonstrating that it maintains benchmark performance in well-specified scenarios. Furthermore, it significantly reduces mean control conflicts by 35.0% in scenarios with unmodeled goals and by 19.1% in those with unmodeled obstacles, while improving the system's usability by 11.2% and 39.5%, respectively. Detailed analysis of human-robot interactions highlights our approach's robustness in tolerating human input noise and adaptability to changes in operator intent.