LAUSR

Autonomous navigation has achieved significant success in structured indoor 2D environments but still encounters major challenges in unstructured outdoor 3D terrains, especially on uneven ground. To address these limitations, we propose a novel navigation framework that integrates continuous traversability estimation and safety-aware planning. The core innovation lies in three aspects: (1) a Bayesian generalized kernel inference method that estimates unobserved point cloud attributes to generate a continuous traversability map; (2) a traversability-aware A* algorithm (TRA-A*), which enhances global path planning by incorporating terrain constraints; and (3) a nonlinear model predictive control (NMPC) that integrates both traversability and uncertainty into the cost function to enable robust motion execution. These components collectively form a safety-focused navigation pipeline capable of stable and adaptive operation in complex terrains. Experimental results from simulations and real-world trials demonstrate that TRA-A* effectively enhances navigation safety and efficiency. Compared to the baseline A*, TRA-A* reduces terrain cost by 42.6% while incurring only a 9.83% increase in path length and a 10.7% increase in the number of nodes. Furthermore, when compared with the state-of-the-art PF-RRT*, it shortens the path length by 13.88%, reduces the number of nodes by 28.70%, and lowers terrain cost by 16.12%, demonstrating superior adaptability. In addition, NMPC significantly improves motion stability, reducing longitudinal acceleration fluctuations by 28.52% and lateral angular velocity fluctuations by 28.99%, thereby enabling more stable and reliable navigation.