Balanced Domain Randomization for Safe Reinforcement Learning

Reinforcement Learning (RL) has enabled autonomous agents to achieve superhuman performance in diverse domains, including games, navigation, and robotic control. Despite these successes, RL agents often struggle with overfitting to specific training environments, which can hinder their adaptability to new or rare scenarios. Domain randomization is a widely used technique designed to enhance policy robustness by training models in a variety of randomized contexts. In traditional domain randomization, however, models tend to prioritize learning from more common domains, often neglecting rare ones. To address this imbalance, we propose Balanced Domain Randomization (BDR) that balances the learning focus based on the rarity of contexts. We assess context rarity in the embedding space using statistical analysis of context vectors and adjust the loss weights for each transition based on this rarity. This ensures that the agent dedicates adequate attention to rare domains. Our experimental results show that BDR efficiently enhances worst-case performance, significantly improving the robustness of RL-based robotic controllers across diverse conditions. This study provides a robust framework for RL agents, reducing risks in uncommon scenarios and ensuring reliable performance in varied environments.