Learning to Adapt: Representation-Based Reinforcement Learning for Multi-Task Skill Transfer

Abstract

Reinforcement learning (RL) has demonstrated remarkable success in learning complex control policies, yet its applicability to real-world robotics remains limited due to sample inefficiency and poor generalization across tasks. This thesis explores representation-based reinforcement learning as a means to address these challenges, with a particular focus on multi-task learning and policy adaptation. We begin by conducting a comprehensive review of state-of-the-art offline RL algorithms, highlighting the evolution from traditional Soft Actor-Critic (SAC) to advanced representation-based methods such as Contrastive Learned Representation Soft Actor-Critic (CTRL-SAC). Motivated by the limitations of existing approaches, we propose a novel multi-skill representation learning framework that builds upon CTRL-SAC, more specifically known as RepMT-SAC, enabling efficient skill transfer by leveraging shared task-independent dynamics. To evaluate the effectiveness of our framework, we conduct extensive empirical studies in order to demonstrate that our approach improves policy adaptation and generalization, reducing the need for task-specific retraining. By combining theoretical insights with rigorous empirical validation, this work contributes to the advancement of reinforcement learning for autonomous control, paving the way for more efficient and versatile robotic systems.