Robust Robot Control for Human-Walking Assistance from Physical Human-Robot Interaction

Safe and effective control of robots to assist humans is complex, especially when it involves close physical interactions with humans. This work presents a hierarchical framework to build robot controllers to assist humans in walking under frailty constraints. We propose to train an RL-based human policy in simulation, to model and synthesize human walking behaviours under external assistance and frailty constraints. It provides a testbed of safe and scalable interactions with humans, and thus allows for iterative fine-tuning of robot behaviours to offer physical assistance robustly. The efficacy of the proposed framework has been evaluated on a dual-arm assistive robot. Experimental results show that the learned walking policy enables humans to leverage random external assistance to generate and stabilize walking motions under frailty constraints. We also demonstrate that the robot controller obtained from interacting with the human is more effective and robust to assist the frail human in walking.