Skill Based Reinforcement Learning

Skill-based reinforcement learning (RL) aims to improve the efficiency and adaptability of RL agents by learning and utilizing reusable, temporally extended skills instead of directly controlling low-level actions. Current research focuses on developing methods for automatically discovering and representing these skills, often leveraging large language models or generative models like diffusion models to create diverse and transferable skill sets from offline data, and employing hierarchical RL architectures for efficient planning and execution. This approach promises significant advancements in sample efficiency and robustness for complex, long-horizon tasks in robotics and other domains, particularly those with sparse rewards or limited training data.