Learning-Based Heuristic Dynamic Path Planning Using a Hop- Aware Graph Neural Network and ConvGRU

<title>Abstract</title> <p>Dynamic path planning in grid environments is challenging because geometric heuristics do not encode changes in obstacle topology, whereas learned heuristics may have limited spatial receptive fields or rely on single-frame observations. This paper presents a learning-based heuristic that combines a hop-aware graph neural network (HopGNN) with a convolutional gated recurrent unit (ConvGRU) and embeds the predicted cost-to-go map in A* search. Successive hop modules construct spatial representations at multiple effective neighborhood scales, and an attention mechanism adaptively fuses their intermediate outputs. ConvGRU integrates a fixed history of grid observations to represent short-term environmental evolution. Experiments on 20 × 20, 30 × 30, and 40 × 40 dynamic grids compare the method with GCN-A*, GAT-A*, D* Lite, and greedy best-first search. On the 40 × 40 evaluation set, HopGNN-A* achieves a success rate of 0.72, compared with 0.66 for GCN-A* and 0.69 for GAT-A*, while reducing mean accumulated expanded nodes by 43.7% and 41.5%, respectively. The hop-count ablation identifies K = 8 as the best tested setting, although the effect of K is non-monotonic. Temporal modeling reduces expanded nodes by 65.4% relative to the no-ConvGRU variant (HopGNN-NoConvGRU). The experiments show higher empirical success rates and more focused search in simulated dynamic grids, together with increased computation time and the need for validation in continuous and real-robot settings.</p>