AnyNav: Visual neuro-symbolic friction learning for off-road navigation

This research presents AnyNav, a vision-based framework that integrates neural networks with symbolic physical models to estimate terrain-vehicle friction and generate physically feasible navigation paths. The system uses an imperative learning paradigm with bilevel optimization to train a friction network in a self-supervised manner, enabling strong generalization across different off-road environments and four-wheeled vehicle platforms. The authors demonstrate successful transfer from simulation to real-world robotic settings, highlighting robustness without asserting full autonomy or commercial deployment.