Learning-based locomotion control fusing multimodal perception for a bipedal humanoid robot

Abstract

The ability of bipedal humanoid robots to walk adaptively on varied terrain is a critical challenge for practical applications, drawing substantial attention from academic and industrial research communities in recent years. Traditional model-based locomotion control methods have high modeling complexity, especially in complex terrain environments, making locomotion stability difficult to ensure. Reinforcement learning offers an end-to-end solution for locomotion control in humanoid robots. This approach typically relies solely on proprioceptive sensing to generate control policies, often resulting in increased robot body collisions during practical applications. Excessive collisions can damage the biped robot hardware, and more critically, the absence of multimodal input, such as vision, limits the robot’s ability to perceive environmental context and adjust its gait trajectory promptly. This lack of multimodal perception also hampers stability and robustness during tasks. In this paper, visual information is added to the locomotion control problem of humanoid robot, and a three-stage multi-objective constraint policy distillation optimization algorithm is innovantly proposed. The expert policies of different terrains to meet the requirements of gait aesthetics are trained through reinforcement learning, and these expert policies are distilled into student through policy distillation. Experimental results demonstrate a significant reduction in collision rates when utilizing a control policy that integrates multimodal perception, especially in challenging terrains like stairs, thresholds, and mixed surfaces. This advancement supports the practical deployment of bipedal humanoid robots.