Title: Contextual Multi-Task Reinforcement Learning for Autonomous Reef Monitoring

Abstract:

While autonomous underwater vehicles (AUVs) hold significant potential for marine ecosystem monitoring, their practical deployment is often hindered by the complexities of controlling these vehicles amidst highly uncertain and non-stationary underwater dynamics. To overcome these obstacles, we utilize a data-driven reinforcement learning strategy designed to adapt to unknown dynamics and shifts in task requirements. Conventional single-task reinforcement learning frequently suffers from overfitting to specific training environments, which restricts the long-term utility of the resulting policies. Consequently, we introduce a contextual multi-task reinforcement learning framework, enabling the development of controllers that can be repurposed across diverse tasks, such as identifying oysters in one reef system and detecting corals in another.

This study assesses the ability of contextual multi-task reinforcement learning to efficiently acquire robust and generalizable control policies for autonomous reef monitoring. We train a single context-dependent policy capable of addressing multiple related monitoring objectives within a simulated reef environment using HoloOcean. Our experimental analysis empirically evaluates these contextual policies in terms of sample efficiency, zero-shot generalization to novel tasks, and resilience against fluctuating water currents. By leveraging multi-task reinforcement learning, we seek to enhance both training efficiency and the reusability of learned policies, thereby advancing more sustainable practices in autonomous reef monitoring.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC