Supervised Momentum Contrastive Learning-Based Coarse-to-Fine Fusion Path Planning

Jin Zhou, David Chieng, Boon Giin Lee, Junkai Ji, Zun Liu, Jianqiang Li

Abstract: Data-driven, learning-based path planning approaches have demonstrated significant advantages over traditional planning algorithms in terms of search efficiency and computational cost. These approaches can learn from large-scale environmental data and plan paths effectively. Currently, mainstream learning-based methods such as Neural RRT and Neural A* can directly output a promising region that connects the start and end points provided by the input map. This region provides a smaller sampling space for sampling-based methods and a more effective heuristic function for search-based methods. However, the paths generated by such learning-based approaches are generally inferior to those of traditional planners. Furthermore, their performance exhibits poor generalization ability, particularly in unseen environments. In this paper, we propose a supervised momentum contrastive learning method to address these challenges. Our method leverages the powerful feature representation and generalization capabilities of contrastive learning to generate a coarse-grained map with promising regions. Subsequently, an any-angle search planner, Theta*, is applied to perform fine-grained path generation on the resulting navigation map. To the best of our knowledge, this is the first work that adopts contrastive learning to solve path planning problems. Results show that the proposed method achieves a better trade-off between solution quality and search efficiency compared to state-of-the-art learning-based planners and traditional planners. Evaluations in out-of-distribution environments and real-world deployments also confirm that the proposed method has a superior generalization performance and satisfactory applicability. Note to Practitioners—This work is motivated by developing a hierarchical path planning system that overcomes the high computational burden and low path quality problems of traditional algorithms and learning-based approaches, respectively. The proposed method reduces the computational cost of traditional planners while enhancing the path quality of learning-based approaches. It demonstrates improved adaptability to unseen environments, addressing a common challenge in practical applications. By employing a hierarchical planning framework, the method significantly improves search efficiency, reduces planning time, and meets real-time requirements for robotic navigation. Specifically, the approach supports cooperative uncrewed aerial/ground vehicle (UAV/UGV) systems. In this framework, the UAV captures real-time global environmental data from an aerial perspective and transmits prior information (e.g., image-based maps) to the ground agent. The UGV leverages this data to generate high-quality navigation paths. This synergy enables the UAV with its broad aerial view and high mobility to efficiently guide the UGV in exploring unknown environments. However, the proposed approach is restricted to processing images of fixed size and low resolution, thereby limiting its capability to analyze the unpredictable and variable aerial view images captured by UAVs across diverse real-world environments. In future research, we will focus on this problem and address it.

External IDs:doi:10.1109/tase.2025.3626557