Enhancing Safety and Explainability of Reinforcement Learning Agents for Environmental Monitoring Tasks

Mitigating pollution in aquatic ecosystems is among the most pressing challenges in environmental sustainabil- ity applications. While effective monitoring and intervention activities are key to safeguarding water quality, protecting biodiversity, and supporting industries (e.g., aquaculture), this is traditionally done by human oper- ators—making the process costly, time-consuming, and often inadequate for capturing timely environmental changes. In this work, we focus on safe, explainable design and deployment of autonomous reinforcement learning (RL) agents for environmental monitoring tasks. In particular, we present our recent contributions to: i) safe RL techniques, ii) Neurosymbolic RL, iii) formal verification of deep RL policies, and iv) designing robust control strategies for safe deployment.