Outdoor Multi-UAV Collaborative Task Planning System Based on Multi-Agent and Improved PPO Algorithm
Runmin Wang, Zhong-Liang Deng · 2025
Aiming at the problem of task collaboration and path planning for multi-agent drones in unknown environments, this paper proposes a multi-agent collaborative planning architecture that integrates a large language model (LLM) and a multimodal (MM) perception network, which is used for executing the multi-agent tasks in outdoor environments. This paper innovatively proposes an improved Proximal Policy Optimization (PPO) algorithm for path planning. Through a structured reward function and a joint strategy sharing mechanism, the path planning among multiple drones were efficiently achieved. The system in this paper innovatively proposes a loop intelligent control process called UTTUT (User Instruction, Task Graph Generation, Task Allocation, UAV Feedback, and Task Graph Updating), as well as a \({P}^3\) (Perception, Planning, Prompt) factor graph model. Theoretically analyzes the interpretability of large model collaborative task planning. The Doubao large language model is introduced as the central control unit, and the Grounding DINO perception model is deployed at the drone end. By combining RGBD, semantic information, and asynchronous scheduling task factors, the perception and positioning of targets in complex scenes are realized. In addition, this paper compares the performances of different large model collaborative frameworks and gets the optimal large model control scheme results. Simulations are carried out on different platforms to verify the superiority of the framework proposed in this paper.