Adversarial Motion Priors Stabilize Humanoid Step Transitions on Procedural Rubble Terrain

Abstract

The deployment of bipedal humanoid robots in unstructured and hazardous environments relies heavily on the development of robust locomotion controllers capable of handling unpredictable disturbances. Traditional control methodologies often struggle to adapt to chaotic terrain, such as disaster rubble, due to their reliance on precise environmental models and rigid kinematic trajectories. Recently, reinforcement learning has emerged as a promising alternative, yet policies trained purely on task-specific rewards often exhibit unnatural, overly aggressive, or energy-inefficient movements that fail when transferred to physical hardware. This paper proposes a novel framework utilizing adversarial motion priors to constrain the reinforcement learning exploration space during humanoid step transitions over procedurally generated rubble terrain. By leveraging a discriminator network trained on a dataset of human motion capture data, the proposed system forces the humanoid to learn stepping strategies that not only accomplish the task of traversing severe terrain but also maintain kinematic realism and dynamic stability. The framework utilizes a robust state-action formulation that seamlessly integrates proprioceptive feedback with exteroceptive elevation maps. Extensive simulation experiments demonstrate that the integration of adversarial motion priors significantly improves both the success rate of traversing dense rubble and the energy efficiency of the locomotion policy compared to baseline reinforcement learning methods. Furthermore, the learned policies exhibit emergent human-like recovery behaviors during critical step transitions, thereby validating the efficacy of adversarial imitation in complex bipedal locomotion tasks.