Safe Reinforcement Learning for Collaborative Robots in Dynamic Human Environments

Collaborative robots trained using Reinforcement Learning (RL) techniques to perform complex tasks have shown promising results in simulations and controlled environments. However, the actions of such autonomous robots are not always predictable in real-world settings. This poses risks of injury to humans, particularly in joint action tasks where human and robot come in close contact. Conventional safe RL models, such as reward shaping and constraint RL, prioritize safety during learning and deployment, but do not guarantee that all actions will lead to safe states. Shielded RL, which seeks to provide safety guarantees, also struggles in joint action tasks such as collaborative assembly and object handover. Improving shielded RL to account for close-contact human-robot tasks presents a potential solution to facilitate the safe application of RL-based robots in the real world. This research aims to develop safe RL models that ensure safe robot motion in human-robot joint action tasks. Additionally, it investigates how these models affect human-robot interaction to identify human-preferred robot motions for collaborative tasks. The outcomes of this research aim to significantly advance the integration of safe collaborative robots in social, healthcare, and industrial environments.