TETR: Traversability Evaluation with Transformers for Autonomous Navigation in 3D Terrain Environments

Autonomous navigation of unmanned ground vehicles(UGVs) in complex three-dimensional terrains necessitates precise evaluation of road traversability, a challenge that remains unresolved due to the difficulty in quantifying passable regions. Existing methodologies often fall short in effectively evaluating traversability in complex environments. In this paper, we present a novel approach for evaluating terrain traversability by enabling the model to implicitly learn principle-based knowledge. The approach utilizes elevation maps obtained from equiped sensors of UGVs, capturing detailed terrain information through a novel transformer-based model, enabling accurate and efficient traversability assessments. We evaluated our method using a comprehensive simulation dataset, which achieves an accuracy of 93.1%, representing a 5.2% improvement over existing baseline method. Furthermore, our approach demonstrated a 16.9% reduction in computational overhead, facilitating real-time implementation.