The article discusses the challenges of using reinforcement learning (RL) in real-world tasks due to dynamics gaps between real and simulated environments. It introduces a new algorithm, H2O+, which aims to address these issues by combining offline and online learning methods. The algorithm is designed to be flexible, accommodating various choices of offline and online learning methods and accounting for the dynamics gaps. The authors demonstrate the superior performance and flexibility of H2O+ over other RL algorithms through simulation and real-world robotics experiments.
Publication date: 25 Sep 2023
Project Page: https://sites.google.com/view/h2oplusauthors/
Paper: https://arxiv.org/pdf/2309.12716
