CREStE: Scalable Mapless Navigation with Internet Scale Priors and Counterfactual Guidance

Fig. 1: CRESTE learns a perceptual encoder 1,2 and reward function 3 to predict structured bird's eye view (BEV) feature and reward maps for navigation.Our model inherits generalization and robustness from visual foundation models and learns expert-aligned rewards using our counterfactual-guided learning framework.We integrate CRESTE in a modular navigation system that uses coarse GPS guidance and rewards to reach navigation goals safely.