Imagine to Ensure Safety in Hierarchical Reinforcement Learning

This paper presents a novel approach to safe exploration in hierarchical reinforcement learning by integrating a learnable world model with a high-level policy that sets subgoals and a low-level policy that uses imagined rollouts to navigate safely. The method demonstrates significant improvements over existing Safe RL baselines in long-horizon navigation and manipulation tasks, achieving higher success rates and better adherence to safety constraints. This advancement is crucial for practitioners as it addresses the limitations of current safe exploration techniques in complex environments, enabling more reliable deployment of RL agents in safety-sensitive applications.