Residual-learning-based landing control with gravity estimation for quadruped robot in low-gravity scenarios

Exploring asteroid surfaces presents unique challenges due to their low-gravity environment and rugged terrain, making landing and mobility difficult for exploration probes. Quadruped robots, with their exceptional mobility, exhibit strong adaptability on rugged planetary surfaces, making them promising tools for future space exploration. However, the complex structure of quadruped robots poses significant challenges for both traditional control methods and learning-based approaches. In this paper, set against the backdrop of previous quadruped robot landing pipelines, we propose a residual strategy that combines the stability of traditional control with the flexibility of reinforcement learning. To be specific, the incorporation of a simple dynamics prior model helps to accelerate the reinforcement learning process by providing structured guidance, while reinforcement learning enhances the overall adaptability of the model, enabling it to flexibly respond to varying environmental conditions. Additionally, the introduction of a gravity estimation module enhances the algorithm’s adaptability in unknown environments. Through ablation experiments and generalization tests, we validate the effectiveness of the proposed method. Furthermore, we successfully apply this method to jumping scenarios, thereby demonstrating its potential in more complex tasks.