Synergizing Multimodal Large Language Models and GRPO-Based Physics-Guided DRL for UAV Swarm Navigation in Dynamic Environments
Abstract
Unmanned aerial vehicle (UAV) swarms are essential for time-sensitive missions such as disaster response, environmental monitoring, and search-and-rescue, where adaptive formation control, collision avoidance, and communication-aware navigation are critical in dynamic environments. However, existing artificial potential field (APF)-based methods suffer from local minima and limited adaptability, while conventional multi-agent deep reinforcement learning approaches rely on single-modal perception, require task-specific retraining for different mission objectives and formations, and generalize poorly to unseen environments. To address these challenges, this paper proposes a physics-informed multimodal large language model (PMLLM) framework that integrates heterogeneous sensory information with high-level textual mission prompts to jointly control UAV acceleration, flight direction, formation adaptation, and communication parameters. The framework incorporates realistic communication delay constraints through carrier sense multiple access with collision avoidance-based medium access control with adaptive contention window and neighbor update interval optimization for communication-aware swarm coordination. A two-stage low-rank adaptation-enabled training strategy is introduced, where retrieval-augmented generation-assisted PMLLM first generates mission-aware improved APF-based supervision for low-rank adaptation-enabled supervised fine-tuning, followed by intent-specific reward generation for group relative policy optimization-based reinforcement learning. Simulation results show that PMLLM reduces the aggregate reward-penalty magnitude by 32.9%–70.3% and average swarm travel time by 12.1%–47.6% for the seen line mission and 13.1%–48.3% for the unseen square mission relative to the considered benchmarks and ablations, while preserving adaptive formation and intruder-avoidance behavior.