TravKAN: Fast and Interpretable Nonlinear Traversability Analysis with Kolmogorov-Arnold Networks

As autonomous robots move into unstructured environments, they must accurately assess terrain traversability to ensure safety. While modern deep learning models and gradient-boosted trees provide high predictive accuracy, they often function as black boxes, offering little insight into why a specific terrain patch is classified as traversable or not. This paper investigates whether Kolmogorov-Arnold Networks (KANs) can provide a more transparent, interpretable, and computationally efficient alternative for traversability estimation.

The authors propose TravKAN, a framework that models traversability as a cell-wise binary classification problem. Unlike standard neural networks that apply nonlinearities at nodes, TravKAN uses KANs to parameterize nonlinear functions directly on edges using learnable B-splines. This architecture allows the model to capture complex nonlinear interactions while remaining compact. A key innovation is the inclusion of novel handcrafted features derived from the reflectivity channel of LiDAR sensors, which provide material-specific information that complements traditional geometric cues.

TravKAN achieves predictive performance that outperforms conventional deep learning models and approaches the high accuracy of XGBoost across diverse datasets, including urban (SemanticKITTI, nuScenes) and off-road (EastPark) environments. Because the model is based on KANs, it supports symbolic extraction—a process where the trained network is converted into a concise, human-readable analytic expression (TravKAN-Lite). This symbolic form provides direct insight into the decision-making process, revealing how specific features contribute to the final traversability score, while maintaining O(1) computational complexity for real-time deployment.

In safety-critical robotics, understanding the rationale behind a navigation decision is as important as the decision itself. By bridging the gap between high-performance nonlinear approximation and symbolic interpretability, TravKAN offers a transparent, efficient, and robust solution for autonomous systems operating in complex, unpredictable terrains.