Learning-based heuristics for risk-aware pathfinding in topographic environments for tactical agent simulation

This work presents a framework for risk-aware pathfinding in topographic terrains, designed to support realistic and tactical route planning in agent-based simulations. The proposed method enhances the A ∗ algorithm by incorporating topographic elevation and risk costs into its path cost estimation function, enabling the computation of paths that reflect tactical constraints in simulation scenarios. To improve the efficiency of heuristic estimations used in path search, we introduce the use of Deep Neural Networks (DNNs) trained to learn correction factors for heuristic functions. These models adjust distance and risk estimates during the search, outperforming conventional heuristic functions, particularly in complex environments. To reduce the computational overhead typically associated with DNN inferences, we use an Iterative Heuristic Map (IHM), a caching and batching mechanism that minimizes redundant model queries by reusing heuristic estimates for spatially related nodes. The framework introduces the notion of View Point Configurations (VPCs), a method to model multiple risk sources in a digital elevation model (DEM) by simulating the field of view from elevated risk points. VPCs enable the representation of dangerous and safe terrain areas and are used to train and evaluate DNN models. The presented work conducted comprehensive experiments in increasingly complex scenarios, including unitary, mixed, and multiple active VPC configurations. It also analyzes the risk exposure of the paths calculated by the proposed framework, comparing them with those generated by the baseline A ∗ algorithm. The results confirm the effectiveness of the framework in reducing computation costs, enhancing risk-aware DNN generalization, and improving tactical pathfinding in topographic terrains.