Leveraging self-supervised learning and mask-based regularization improves traversability prediction in off-road environments.