Bipedal Navigation and Locomotion Control Algorithms for Humanoid Robots

Abstract

With their anthropomorphic design, humanoid robots are naturally adapted to human environments; however, high degrees of freedom, severe nonlinear dynamics, and small stability margins make autonomous locomotion a major bottleneck for practical applications. This paper surveys the key algorithms for humanoid robot locomotion, navigation, and motion control. In legged navigation, this survey traces the development from simplified-model gait generation to search-based footstep planning, and finally to online dynamic replanning and, ultimately, a hierarchical "Footstep Planning to Model Predictive Control to Whole-Body Control" framework. In motion control, this paper reviews two primary approaches: model-based optimization and reinforcement learning. The former has gone from simple models to real-time model predictive control and whole-body multi-contact optimization, and the latter has developed from proprioceptive blind walking to terrain adaptation through vision and synthesis of anthropomorphic motion. Next, a comparative analysis of these two paradigms on three aspects: stability, real-time performance and environmental adaptability, reveals that the "model-based reference with learning-based residual" hybrid scheme has emerged as a prominent trend. Finally, open challenges are mentioned in a systematic fashion, such as the interpretability and safety of learned policies, long-horizon robust locomotion in unstructured environments and whole-body integration of locomotion and manipulation.