Domain Adaptation for Outdoor Robot Traversability Estimation from RGB data with Safety-Preserving Loss

Being able to estimate the traversability of the area surrounding a mobile\nrobot is a fundamental task in the design of a navigation algorithm. However,\nthe task is often complex, since it requires evaluating distances from\nobstacles, type and slope of terrain, and dealing with non-obvious\ndiscontinuities in detected distances due to perspective. In this paper, we\npresent an approach based on deep learning to estimate and anticipate the\ntraversing score of different routes in the field of view of an on-board RGB\ncamera. The backbone of the proposed model is based on a state-of-the-art deep\nsegmentation model, which is fine-tuned on the task of predicting route\ntraversability. We then enhance the model's capabilities by a) addressing\ndomain shifts through gradient-reversal unsupervised adaptation, and b)\naccounting for the specific safety requirements of a mobile robot, by\nencouraging the model to err on the safe side, i.e., penalizing errors that\nwould cause collisions with obstacles more than those that would cause the\nrobot to stop in advance. Experimental results show that our approach is able\nto satisfactorily identify traversable areas and to generalize to unseen\nlocations.\n