Curriculum-Based Deep Reinforcement Learning for Robust Biped Push Recovery

Push recovery - the ability to regain balance after unexpected external disturbances - is a key capability for deploying biped robots in real-world environments. Existing approaches often rely on simplified template models, which can be overly conservative and fail to capture realistic recovery behaviors under disturbances that vary in location, magnitude, and duration. This paper presents a push-recovery framework that combines a sequential curriculum learning strategy with massively parallel deep reinforcement learning. By training the policy on progressively more challenging disturbance scenarios, the proposed method enables robust recovery from pushes applied at different body locations with varying intensity and duration. Simulation results demonstrate strong performance, including demanding dual-disturbance cases with simultaneous external forces.