Using Adversarial Reinforcement Learning to Improve the Resilience of Human-Robot Collaboration in Industrial Assembly

The paper proposes a novel approach to enhance the resilience of mutual collaborative activity between humans and robots in industrial assembly tasks. The approach exploits Adversarial Reinforcement Learning (ARL) to enable a robot to learn an assembly policy that is robust against human mistakes. The adversary can represent various sources of uncertainty or disturbance in the environment. By learning from adversarial feedback, the agent can improve its performance and adaptability in challenging scenarios. The paper applies ARL to the execution of the assembly task sequence. The robot acts as one agent and learns how to assist the human partner during the assembly. The agent simulating the human partner acts as the adversary and deliberately introduces mistakes during the assembly process. The robot also learns how to cope with different levels of human competence and cooperation by adjusting its own behaviour accordingly. The paper evaluates the proposed approach through experiments reproducing complex assembly sequences and compares it with baseline methods that use conventional optimization algorithms. The results show that ARL does not outperforms conventional optimization algorithms in terms of task completion time but guarantee robustness against human mistakes. The paper also discusses the implications for human-robot collaboration and suggests future directions for research.