WayFAST: Navigation With Predictive Traversability in the Field

We present a self-supervised approach for learning to predict traversable paths wheeled mobile robots that require good traction navigate. Our algorithm, termed WayFAST (Waypoint Free Autonomous Systems Traversability), uses RGB and depth data, along with navigation experience, autonomously generate in outdoor unstructured environments. key inspiration is can be estimated rolling using kinodynamic models. Using estimates provided by an online receding horizon estimator, we are able train traversability prediction neural network manner, without requiring heuristics utilized previous methods. demonstrate the effectiveness of through extensive field testing varying environments, ranging from sandy dry beaches forest canopies snow covered grass fields. results clearly learn avoid geometric obstacles as well untraversable terrain, such snow, which would difficult sensors provide only LiDAR. Furthermore, show our training pipeline based on more data-efficient than other heuristic-based