Predictive Safety Curricula for Robust Legged Locomotion
This study addresses the insufficient deployment robustness of legged robots arising from training distributions that fail to cover rare, safety-critical scenarios. To overcome this limitation, we propose a predictive safety curriculum framework that innovatively introduces a distributed safety critic to estimate future safety costs and guide curriculum generation, dynamically optimizing terrain and stochastic event distributions. By integrating reinforcement learning with distributed value estimation, this approach transcends conventional methods restricted to task difficulty adjustment or advantage-based replay. Extensive evaluations in both simulation and on a physical ANYmal-D platform demonstrate significant reductions in collision rates, achieving a 63% decrease in shank collisions. Furthermore, when deployed on a production-grade stair-climbing system, the proposed method completely eliminates all observed collision incidents, confirming its effectiveness in enhancing real-world safety for legged locomotion.