Human-robot collaboration simulation environment with reinforcement learning rescheduling agent

Human-robot collaboration (HRC) is a growing research area where robots and humans work together to accom- plish tasks that require both cognitive flexibility and precision. This thesis focuses on optimizing task scheduling in a collaborative setting where a human operator and a cobot (collaborative robot) assemble mosaic tiles on a 4x5 grid. The objective is to minimize both the human operator’s stress and the overall time to complete the as- sembly. Real-world experiments were conducted with various operators wearing Pupil Core eye-tracking glasses to measure physiological stress while performing tasks, and a robot equipped with a vision system. Tasks were pre-assigned based on reachability constraints, and both fixed and rescheduled task experiments were conducted. To explore the potential for optimizing task allocation, a simulation environment was developed using the OpenAI Gymnasium framework. This environment was designed to emulate real-world conditions by sampling task completion times and stress levels from the collected data. Within this simulated environment, a value-based reinforcement learning (RL) algorithm was implemented to develop optimal task scheduling strategies that bal- ance the goals of minimizing stress and reducing the overall time required to complete the mosaic assembly tasks. Given the complexity of this multi-objective optimization problem, the value-based RL approach allowed the agent to learn and estimate the value of different task scheduling decisions by sampling and evaluating a batch of legitimate task assignments for both the human operator and the robot. By refining task allocation over time, the agent demonstrated the potential to optimize task schedules in this dynamic and sequential decision-making environment.