FB-MEBE: New Algorithm for Sim2Real Zero-Shot RL in Quadrupedal Robots

A novel algorithm named FB-MEBE has been developed to tackle the difficulties associated with online zero-shot reinforcement learning (RL) for controlling quadrupedal robots in real-world environments. This research builds on the Forward-Backward (FB) algorithm and is elaborated in a paper available on arXiv (2603.25464v2). The authors note that undirected exploration in online zero-shot RL results in low-diversity data, leading to suboptimal performance and making policies unsuitable for direct application on hardware. To address this issue, FB-MEBE integrates an unsupervised behavior exploration method with a regularization critic, enhancing exploration to improve the pretraining dataset's quality and the effectiveness of policies across various tasks. This study is crucial for advancing robotics and AI, particularly in deploying RL agents in real-world scenarios without prior task knowledge.