Vision-Based Autonomous Navigation for Unstructured Planetary Analog Terrain

Autonomous traversal of unstructured, GNSS-denied terrain remains a central challenge for planetary rovers. We present a vision-centric navigation framework that fuses stereo depth estimation, visual-inertial odometry, and semantic terrain segmentation to construct a real-time traversability map. A sampling-based local planner generates dynamically feasible trajectories, while a confidence-aware recovery policy mitigates localisation drift over long traverses. Field trials on Mars-analog terrain demonstrate robust obstacle avoidance and reliable waypoint navigation without external positioning. The approach generalises across varied lighting and surface conditions, offering a practical autonomy stack for student-built competition rovers.