EVORA: Deep Evidential Traversability Learning for Risk-Aware Off-Road Autonomy

Traversing terrain with good traction is crucial for achieving fast off-roadnavigation. Instead of manually designing costs based on terrain features,existing methods learn terrain properties directly from data viaself-supervision to automatically penalize trajectories moving throughundesirable terrain, but challenges remain to properly quantify and mitigatethe risk due to uncertainty in learned models. To this end, this work proposesa unified framework to learn uncertainty-aware traction model and planrisk-aware trajectories. For uncertainty quantification, we efficiently modelboth aleatoric and epistemic uncertainty by learning discrete tractiondistributions and probability densities of the traction predictor's latentfeatures. Leveraging evidential deep learning, we parameterize Dirichletdistributions with the network outputs and propose a novel uncertainty-awaresquared Earth Mover's distance loss with a closed-form expression that improveslearning accuracy and navigation performance. For risk-aware navigation, theproposed planner simulates state trajectories with the worst-case expectedtraction to handle aleatoric uncertainty, and penalizes trajectories movingthrough terrain with high epistemic uncertainty. Our approach is extensivelyvalidated in simulation and on wheeled and quadruped robots, showing improvednavigation performance compared to methods that assume no slip, assume theexpected traction, or optimize for the worst-case expected cost.