A reinforcement learning framework for energy-optimal UAV path planning in wind fields

Energy-efficient path planning for unmanned aerial vehicles (UAVs) in low-altitude environments with dense obstacles and uncertain wind disturbances remains a fundamental challenge in autonomous navigation. This paper presents a novel deep reinforcement learning framework called Probabilistic Convolutional Q-Network (PQN). The framework integrates three key components: a physics-based energy consumption model, probabilistic roadmap (PRM) planning, and a convolutional neural network for wind field perception. The energy model characterizes UAV propulsion power under wind conditions, while the PRM generates sparse collision-free graphs to reduce the planning search space. A convolutional neural network encoder extracts global wind field features and integrates them into the deep Q-network for wind-aware path planning. Experiments demonstrate that PQN consistently outperforms baseline methods across diverse scenarios. Compared to conventional deep Q-network approaches, PQN achieves 28.1 % reduction in energy consumption, 16.0 % shorter path length, and 74.2 % faster planning time, which highlights the practical potential of PQN for real-world UAV navigation applications.