The study introduces a new method for estimating terrain traversability, crucial for autonomous systems in outdoor environments. The method employs a visual self-supervised learning approach for traversability prediction, utilizing contrastive representation learning with human driving data and instance-based segmentation masks. This technique outperforms other methods in predicting traversability for both on- and off-trail driving scenarios. The researchers demonstrate the technique’s compatibility with a model-predictive controller and its unprecedented performance for generalization to new environments.
Publication date: 29 Dec 2023
Project Page: https://sites.google.com/view/visual-traversability-learning
Paper: https://arxiv.org/pdf/2312.16016
