Detection-Guided Terrain-Aware Autonomous Navigation and Active Target Detection for Quadruped Robots

Abstract

Autonomous navigation of quadruped robots in unstructured, cluttered terrain while simultaneously searching for targets of interest remains challenging because conventional geometric planners ignore the locomotion constraints imposed by slopes, steps, and rough ground, and because passive perception pipelines detect targets only when they happen to fall within a fixed forward field of view. This paper presents DTQ-Nav, a detection-guided, terrain-aware navigation framework that couples a compact convolutional target detector with a 2.5D traversability-aware global planner, a reactive local controller subject to a static gait-stability constraint, and an active-perception policy that steers the sensor toward unconfirmed targets. The perception module is trained and evaluated on a procedurally generated egocentric detection dataset of 6,600 scenes; it attains an average precision of 0.941 for target presence, a receiver-operating-characteristic area under the curve of 0.896, an F1-score of 0.866, and a mean bearing error of 1.55°. The navigation system is assessed over 150 paired Monte-Carlo episodes against four representative baselines (geometric A*, RRT*, a reactive DWA/APF controller, and an information-gain explorer). DTQ-Nav reaches every goal (success rate 1.00) with zero collisions, raises the target detection rate to 0.721 (versus 0.285 for passive reactive navigation) and reduces time-to-first-detection by an order of magnitude, while lowering locomotion energy by approximately 10.9% through terrain-aware routing. Ablation of components confirms that terrain-awareness governs energy efficiency, active perception governs detection, and reactive replanning governs safety. The framework is directly applicable to search-and-rescue, infrastructure inspection, and planetary exploration using legged platforms.