Neuromechanical Model-free Epistemic Risk Guided Exploration (NeuroMERGE) for Safe Autonomy in Human-Robot Interaction

Optimal human-robot interaction (HRI) necessi-tates the ability to track and compensate nonlinear neuromus-cular and biomechanical dynamics that are challenging to iden-tify online during movement. Model-free reinforcement learning approaches are well-suited to identifying such system dynamics through stochastic exploration and subsequent exploitation of learned low-dimensional probabilistic models to maximize reward. However, achieving safe and efficient stochastic explo-ration in HRI environments is an unsolved challenge. This work presents the development and experimental validation of a Neu-romechanical Model-Free Epistemic Risk-Guided Exploration (NeuroMERGE) algorithm for stochastic iterative identification of HRI dynamics, a novel approach which integrates a mea-surement model of neuromechanical impedances to dynamically constrain the exploration-exploitation tradeoff. We validate NeuroMERGE in the control of a simulated cart-pole system as well as in a soft robotic hand exoskeleton in a case study with three participants. The results demonstrate safe and efficient convergence to stable control policies, achieving performance competitive with model- and learning-based control schemes.