Efficient deployment of multiple jumping robots in uneven terrains using deep reinforcement learning

Biologically-inspired jumping robots are capable of leaping over or onto obstacles, showcasing remarkable environmental adaptability. However, path planning for efficient deployment of multiple jumping robots remains a difficult challenge in uneven three-dimensional terrains. In this work, we present a locust-inspired jumping robot (JumpBot) with multiple locomotion modes (crawling, turning and jumping), and propose a multi-robot coordination algorithm (MCA) using deep reinforcement learning. MCA employs a centralized training framework with decentralized execution to enhance training efficiency. Additionally, we integrate long short-term memory (LSTM) networks into the training framework, which improves the ability of policy networks to process critical features for effective robot collaboration. For multi-target autonomous deployment tasks, we developed a simulation platform with experimental scenarios of different sizes and randomly placed obstacles. Simulation results demonstrate that JumpBot effectively combines both jumping and crawling modes, reducing the average path cost by 22.9 % compared to crawling alone. Moreover, our algorithm achieved an 81.2 %±2.39 % success rate, outperforming typical benchmark algorithms. Finally, we completed the deployment task of multiple jumping robots in a real-world environment for the first time, providing a novel approach to intelligent decision-making and collaboration for terrestrial robots.