Meta-Reinforcement Learning for Enhanced Generalization in Robotics

This thesis aims to investigate the use of meta-reinforcement learning (meta-RL) to enhance sample efficiency in robotic systems, enabling agents to adapt to unseen tasks with fewer interactions. By employing contextual Markov Decision Processes (cMDPs) to formalize task variability, the study offers a structured way to describe tasks within the meta-RL framework. The primary objective is to achieve comparable generalization performance while reducing the number of interactions needed for adaptation. Traditional reinforcement learning methods often require large amounts of data to adapt to new tasks; however, by leveraging prior experiences, meta-RL has the potential to improve learning efficiency and adapt more quickly to novel situations. To further refine exploration strategies, the research plans to incorporate adversarial training methods. An adversary will generate increasingly challenging scenarios, pushing the agent into difficult and less-explored states. This approach is expected to drive the agent to discover more effective strategies, enhancing exploration and ultimately improving learning efficiency.