Reinforcement Learning Subnetwork Discovery for AUV Control

A new arXiv preprint (2604.21640) investigates task-specific subnetwork discovery in reinforcement learning for autonomous underwater vehicle (AUV) navigation. The research addresses the opacity of multi-task RL policies, which perform well in simulations but lack transparency for real-world deployment. By analyzing the internal structure of pretrained policies, the authors aim to uncover task-specific specializations and improve interpretability, trust, and safety. The work targets challenges in dynamic, uncertain underwater environments with limited sensing, where classical controllers fail. The study focuses on enabling robust, generalizable, and explainable control policies for long-term monitoring.