The paper, by Qizhen Wu, Kexin Liu, Lei Chen, and Jinhu Lü, addresses multi-agent UAV pathfinding in stochastic environments. Traditional planning methods need repeated iterations as the environment changes, which becomes computationally expensive, especially for decentralized agents without a centralized planner. Reinforcement learning offers generalization but normally requires enormous agent-environment interactions during training.
The authors propose a centralized-training, decentralized-execution method based on multi-agent reinforcement learning, improved with ideas from model predictive control. During online operation, agents communicate only with a centralized planner to make decentralized decisions. To respect communication constraints, each agent plans using an extended observation that combines neighboring-agent information through a distance-weighted mean-field approach.
Inspired by the rolling optimization of model predictive control, the method conducts multi-step value convergence to improve training efficiency and reduce expensive interactions. The authors report experiments including comparisons, ablations, and real-robot studies that validate the effectiveness and generalization of the approach. As a single arXiv preprint, these claims are not independently verified.