Instance-specific heuristic learning for scalable UAV path planning

Abstract

Efficient and scalable path planning is critical for autonomous UAVs navigating complex, obstacle-dense environments. Traditional heuristic search algorithms like A∗ and Focal Search often face challenges with scalability and adaptability in such scenarios. We present a framework that leverages Transformer-based heuristic learning to predict Path Probability Maps (PPM), which are probabilistic grids that indicate the likelihood of each cell being part of an optimal path from start to goal. This significantly enhances search efficiency and solution quality by guiding the search towards high-probability regions. Trained on diverse motion planning datasets and tailored for UAV-specific challenges, our method reduces computational overhead while maintaining near-optimal path quality. Empirical results demonstrate the framework’s effectiveness, solving over 50% of scenarios optimally and reducing node expansions by a factor of 2. Additionally, the framework exhibits robust scalability across varying instance sizes, highlighting the potential of instance-dependent heuristic learning to transform UAV path planning for real-time applications.