Scalable and generalizable path planning for robotic navigation using transformer-based heuristic learning

Abstract

Efficient and scalable path planning is a critical challenge for autonomous robotic systems, particularly in complex real-world scenarios. Traditional heuristic search algorithms like A* often struggle with scalability and adaptability in such environments. To address these limitations, we improve a search framework that integrates learned, instance-specific heuristics with conventional pathfinding techniques. Leveraging autoencoder transformer networks, we predict two key heuristic functions—Correction Factor (CF) and Path Probability Map (PPM)—trained on diverse datasets—the Motion Planning (MP) and Tiled-MP datasets—to cover a wide range of path planning scenarios. When integrated with Weighted A* (WA*) algorithm, this approach optimally solves 88% of MP instances, with paths averaging less than 0.7% longer than optimal, and requiring nearly five times fewer node expansions. The framework demonstrates the advantages of heuristic learning in handling larger path planning problems, with inference time accounting for just 10% of the total search duration. It solves nearly half of the most complex instances optimally, showcasing strong scalability for real-time robotics applications. The framework performs well in unseen environments, solving over 25% of new problems perfectly, finding near-optimal solutions with paths less than 7% longer than optimal, and requiring fewer than two-thirds of the typical expansions. Our framework outperforms learnable planners in both scalability and generalization.