A Reinforcement Learning Framework for Real-Time Multi-Agent Manipulation Control of Collaborative Robots

The increasing demand for robots across various industries has led to the development of sophisticated robotic systems requiring precise and efficient control. However, achieving true autonomy in dynamic environments remains a significant challenge. This thesis presents a comprehensive framework that bridges the gap between human expertise and robotic autonomy, enabling efficient and safe robot control, particularly in custom production environments. By leveraging deep learning and robotics, we address key challenges such as accurate robot modeling and model-based control. Our contributions include a deep learning-based approach for dynamic robot modeling, enhancing control and simulation accuracy for serial and parallel robots. Additionally, we integrate model-based control techniques for robots with closed control architectures, analyzing their tracking performance and convergence behavior. To facilitate seamless human-robot interaction, we design an intuitive augmented reality interface that enables high-quality demonstration data collection and supports fast telemanipulation, improving real-time control and safety in tasks requiring precision and speed. This dissertation contributes to robotics, artificial intelligence, and human-robot interaction by developing a comprehensive human-robot interaction framework that addresses critical challenges in robot control and interaction. Future research directions include gaze-based intention recognition and prediction, as well as applying AI techniques to enhance human-robot collaboration.