WOMBET: World Model-Based Experience Transfer for Robust and Sample-efficient Reinforcement Learning

The article introduces WOMBET, a novel framework for experience transfer in reinforcement learning that combines world modeling with uncertainty-penalized planning to generate and utilize prior data effectively. WOMBET enhances sample efficiency and performance in continuous control tasks by filtering trajectories based on return and epistemic uncertainty, and it employs adaptive sampling for online fine-tuning. This approach addresses the challenges of data collection in robotics, making it significant for practitioners aiming to improve the robustness and efficiency of RL systems.