Optimization of multi-agent smoke coverage strategies in dynamic environments based on deep reinforcement learning

Cooperative smoke coverage using multiple unmanned aerial vehicles (UAVs) is critical for dynamic regional obscuration in emergency response scenarios, yet remains challenging due to wind disturbances, turbulent diffusion, and complex multi-agent dynamics. Existing coverage control methods typically target static regions and struggle to address non-stationary smoke evolution or persistent coverage maintenance relying on real-time environmental perception. To overcome these limitations, this paper proposes a wind-aware hierarchical smoke coverage architecture (WA-HSCA), a deep reinforcement learning approach designed for multi-UAV active scalar-field manipulation. By formulating the deployment as a finite-horizon control problem, we construct a simulation environment integrating pseudo-random wind fields, Lagrangian smoke diffusion, and non-holonomic UAV dynamics. Within this architecture, a wind-aware proximal policy optimization (PPO) network serves as the macro-level planner, fusing spatially distributed sensor data to generate high-level navigational waypoints. Operating seamlessly at the micro-level, an artificial potential field (APF) controller executes localized collision avoidance based on spatial proximity. To effectively train this hierarchical system, a multi-objective reward mechanism is designed to promote upwind guidance, inter-agent dispersion, and continuous coverage maintenance. Simulation results demonstrate that WA-HSCA leverages environmental sensing to induce cooperative behaviors, ensuring rapid and stable smoke coverage against strong, non-stationary wind disturbances.