Path planning of locust-inspired jumping robots in obstacle-dense environments using curriculum reinforcement learning

Biologically-inspired jumping robots have demonstrated remarkable adaptability in complex environments, making them increasingly valuable across various fields. However, effective path planning in obstacle-dense environments for large-scale jumping robots remains a significant challenge. Inspired by independent decision-making in the efficient collaborative behavior of locust swarms, we propose a two-stage curriculum reinforcement learning (TS-CRL) framework for locust-inspired jumping robots. This framework enables individual robots to autonomously determine actions based on local environmental observations during group crossing tasks. TS-CRL incorporates a population-invariant encoder with an attention mechanism, allowing it to efficiently handle an increased number of training robots. Moreover, it employs an actor-critic network architecture based on Kolmogorov-Arnold networks to enhance training performance. To further improve the training efficiency, we divided the policy training process into two stages with gradually increasing environmental complexity. The effectiveness and scalability of TS-CRL were validated through a locust-inspired jumping robot platform in challenging simulation scenarios. Notably, TS-CRL can generate efficient, collision-free paths to guide multiple jumping robots. Compared with typical reinforcement learning algorithms, TS-CRL reduced the average path cost by 13.7% and markedly improved the success rate of robots in reaching the target areas. Finally, we constructed a multi-robot system consisting of locust-inspired jumping robots for experiments in the real world.