Safety-Aware Adaptive Reinforcement Learning for Mobile Assistive-Care Robots

We introduce a safety-aware adaptive reinforcement learning (RL) approach for autonomous robots operating in dynamic environments, with a focus on assistive-care applications. To that end, we combine online planning, model-free RL, and awareness of safety constraints within a self-adaptive system architecture. Our approach also incorporates a prediction model to forecast future states, and a safe-action selection mechanism. The evaluation of the approach for a simulated assistive-care robot navigating a large open-plan kitchen with changing obstacle positions shows that it reduces collision rates compared to a baseline RL method, at the expense of a slightly lower reward.