Robust Locomotion Control for Lunar Exploration on Deformable Terrain

With increasing interest in planetary environmental studies, lunar exploration has become a key focus of scientific research. However, the soft soil and complex unstructured terrain on the lunar surface limit the mobility of traditional wheeled lunar rovers. In contrast, legged robots have attracted widespread attention in recent years due to their flexible motion capabilities. Using legged robots to perform challenging lunar exploration missions can expand the scope and depth of scientific research, which is of great significance for the construction of lunar bases and resource exploration. In this paper, we built a lunar simulation environment, proposed a linear representation based on the neural Koopman operator to learn the nonlinear foot-ground contact dynamics model, and used this model to improve the model-based motion controller in deformable terrain. The experimental results in the lunar simulation environment show that the Koopman-based model predictive control incorporating the learned foot-ground contact model significantly improves the robot's velocity tracking performance and body stability in deformable terrain.