Robot Adaptability to People with Different Capabilities through Reinforcement Learning

Robots are increasingly envisioned to offer sensible assistance across diverse settings. This thesis primarily centers on human-robot collaboration involving individuals with potential physical limitations. We aim to enable robots to perceive and adapt to human capabilities acutely, thereby offering personalized assistance. We present a framework for the online estimation of human capabilities based on Reinforcement Learning and Bayesian inference, including a series of estimation strategies and capability-guided exploration algorithms.

During Reinforcement Learning, the agent accumulates evidence that permits the continuous updating of human capability estimates via Bayesian inference. These estimates are pivotal in optimizing robot behavior. By modeling human capabilities as preconditions for specific robot actions, the agent can either deactivate the action or refer to the most suitable actions derived from a pre-training policy aligned with capability estimates.

We present three human-robot collaboration experiments to demonstrate and validate our framework: two are operated within a simulation environment, and one is a real-world experiment. The results show that our methodology accurately estimated human capabilities in static and dynamic environments. Furthermore, the robot exhibited faster adaptation and enhanced performance when human capabilities were incorporated into the learning process.