Active 3D Object Localization of FLS-Detected Targets via Deep Reinforcement Learning

Accurate three-dimensional target localization remains a critical challenge in underwater robotics, especially when relying on Forward-Looking Sonar (FLS) for sensing. While FLS provides reliable range and azimuth information, its inherent ambiguity in elevation angle often hampers precise localization. In this initial study, we explore the use of a Particle Filter-based tracking framework coupled with a Deep Reinforcement Learning (DRL) agent that actively guides an Autonomous Underwater Vehicle (AUV) to informative vantage points. Our Particle Filter approach represents targets via probabilistic anchors, sampling from sonar bounding box data to handle the elevation uncertainty. Instead of passively accepting observations, the DRL agent generates control commands that strategically reposition the AUV to reduce sensor ambiguity. We validate our method in a custom simulation environment implementing kinematics and sensor field-of-view and a realistic noise model for bounding box detections. The preliminary results indicate that a DRL-based approach represents a feasible solution for real-time 3D target localization in underwater environments.