DGT-Map: Directional Global Traversability Mapping Utilizing Multi-Task Learning for Heterogeneous Vehicles
Abstract: Off-road traversability is direction-dependent and influenced by vehicle dynamics, yet most traversability maps assign a single isotropic cost per location and cannot generalize across vehicles. This work introduces DGT-Map, a self-supervised framework for learning directional, global traversability costmaps from RGB-D observations. The method predicts a locomotion-derived signal for terrain patches using a convolutional neural network trained on fused color and geometric features. By utilizing multi-task learning across multiple vehicles, the model learns a shared terrain representation while still capturing vehicle specific differences.. At inference, DGT-Map generates heading-indexed costmaps that integrate into a Hybrid A* planning framework. The system is evaluated on challenging terrains, including slopes that are traversable downhill but not uphill and ridge obstacles feasible for some vehicles but not others. Across tasks and vehicles, DGT-Map improves navigation success by 30-40% over geometric, binary, and learned direction-agnostic baselines.
Advisor: Professor Constantinos Chamzas
Committee: Professor Jing Xiao, Professor Nitin Sanket