Multi-task learning for multi-goal path planning with path-cost estimation and heuristic-guided RRT sampling

Multi-goal path planning (MGPP) requires identifying a closed, collision-free trajectory that visits a set of goals in an optimal sequence. This is typically formulated as a Physical Traveling Salesman Problem, where the quality of the visiting order depends on an undirected complete graph with precise edge weights representing inter-vertex path lengths. However, calculating these weights efficiently in obstacle-rich environments remains a challenge, often leading to a trade-off between sequencing optimality and computational overhead. To address this, we propose a multi-task learning model termed Prior Information Extraction (PIE). The PIE model concurrently estimates inter-vertex path lengths for graph construction and generates “promising regions” and “search guidelines” to serve as low-level heuristics. These outputs are integrated into a sampling-based framework, PIE-MGPP, which bridges high-level sequencing with low-level motion planning. Specifically, PIE-MGPP utilizes the model-generated weights to determine a prioritized visiting order, while the predicted regions and guidelines focus the RRT-based search toward high-probability configuration spaces. Extensive numerical experiments based on training-source and unseen maps demonstrate that PIE-MGPP achieved high planning success rates and shorter paths among successful cases, while using fewer samples than evaluated baselines.