Learning-Based Reward Functions for Improved Autonomous Navigation in Confined Terrains

Efficient autonomous navigation in unknown and confined terrains remains challenging, but crucial for robust robotic exploration and inspection tasks. Current approaches struggle due to the inherent difficulty in navigating unfamiliar environments; a capability humans and animals intuitively develop through experience. While learning-based methods can mimic such intuition, they usually require extensive labeled data, resulting in poor sample efficiency, particularly under sparse supervision. This project proposes a self-supervised, learning-based approach that leverages environmental scans to generate effective supervisory signals.