Lunar Leap Robot: 3M Architecture–Enhanced Deep Reinforcement Learning Method for Quadruped Robot Jumping in Low-Gravity Environment

Abstract

Legged robots offer advantages such as rapid mobility, adept obstacle-surmounting capabilities, and long mission life in lunar exploration missions. The low-gravity environment on the Moon enhances these benefits, particularly in enabling efficient jumps. However, challenges arise from the complexity of the jumping motion and the difficulty in maintaining stability. This paper introduces an innovative algorithm that integrates deep reinforcement learning with a main point trajectory generator, thus providing a reference for training with minimal reliance on human intuition and prior knowledge. Additionally, fine-grained policy optimization is achieved through a multistage reward structure based on the decomposition of the jumping process. Further, the concept of multitask experience-sharing is proposed to facilitate efficient learning across tasks involving plain terrain jumping and overcoming large obstacles. Simulation results demonstrate the effectiveness of the proposed algorithm in achieving precise and stable jumps, reaching heights approximately five times the robot’s height and distances over five times its body length under lunar gravity. Moreover, the robot exhibits agile strategies, successfully overcoming platforms with a height of 2.5 times its body.