Abstract
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, but challenges remain to properly quantify and mitigate risksdue to uncertainties in learned models. This work efficiently quantifies bothaleatoric and epistemic uncertainties 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. The proposed risk-aware plannersimulates state trajectories with the worst-case expected traction to handlealeatoric uncertainty, and penalizes trajectories moving through terrain withhigh epistemic uncertainty. Our approach is extensively validated in simulationand on wheeled and quadruped robots, showing improved navigation performancecompared to methods that assume no slip, assume the expected traction, oroptimize for the worst-case expected cost.