HexaWorld: bridging AI and marine policy through explainability

Reliable marine biodiversity monitoring is essential to support climate adaptation, spatial planning, and the management of marine protected areas. However, high operational costs and limited coverage often hinder large-scale data collection. Artificial Intelligence (AI), particularly through autonomous underwater vehicles (AUVs), offers promising solutions to optimize these efforts. Yet, their integration into real-world decision-making remains limited, as it is often difficult to access the reasoning behind the strategies learned by AI systems. This makes it challenging for non-technical stakeholders to fully engage with and trust these tools, limiting their contribution to transparent and inclusive governance. We present HexaWorld, a simulation framework designed to develop explainable reinforcement learning (RL) strategies for marine biodiversity monitoring. A key feature is the definition of the reward function, which links agent behavior to ecological and operational objectives, such as maximizing biodiversity discovery, avoiding redundant paths, and ensuring safe return to base, making the learned strategies interpretable and actionable. HexaWorld supports both square and hexagonal grid environments and has been tested on three simulated marine habitats: temperate, tropical, and deep- sea ecosystems. We evaluated agent performance based on biodiversity coverage and exploration efficiency. Results show that hexagonal grids improve exploration in complex, obstacle-rich environments, particularly in deep habitats where navigation is more demanding. By providing a flexible and explainable simulation tool, HexaWorld helps develop AI-based strategies that are understandable and ready to inform evidence-based, interactive decision-making in marine conservation and climate governance.