Deep Learning-Enhanced Agoraphilic Navigation for Uneven Terrain and Dynamic Obstacles
Year: 2026
Type: article
Abstract: Autonomous navigation in uneven and dynamic terrains requires reliable estimation of traversable free space and the ability to anticipate how moving obstacles will influence future motion. Traditional Agoraphilic* navigation provides free-space-driven steering but depends on cell-based geometric analysis, which is computationally intensive and sensitive to discretization. This article introduces an enhanced Agoraphilic* framework that integrates deep learning-based free-space estimation, automated data labeling, and predictive dynamic-obstacle reasoning to improve performance in real-world environments. A repurposed PointNet architecture is employed as a regressor that directly analyzes sectorwise point clouds to infer stability-aware free space without grid-based preprocessing. A fully automated labeling procedure is developed to generate sector-level free-space labels from terrain geometry and robot constraints, eliminating manual annotation and enabling scalable dataset creation. To support navigation in dynamic environments, a complementary prediction module estimates future free-space constriction by decomposing obstacle motion and evaluating a time-to-free-space-change. These static and dynamic free-space cues are fused within the steering formulation to produce efficient, geometry-aware, and responsive navigation commands. Real-world experiments in uneven terrains, including scenarios with moving obstacles, demonstrate that the proposed approach significantly reduces computation time, improves robustness to terrain irregularities, and enables reliable navigation in both static and dynamic settings.
Institution: Federation University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204921652