Embodied AI for aquaculture monitoring: From stationary sensing to mobile robotic observation

Abstract

Aquaculture currently accounts for over 57% of global fisheries production and is projected to meet more than 60% of seafood demand by 2030. However, its efficiency and sustainability are hindered by challenges such as inadequate environmental monitoring and stock assessment. This review investigates the application of embodied artificial intelligence (AI) in mobile robotic monitoring for intensive aquaculture. It contrasts this approach with conventional AI, which depends on fixed data sources, highlighting how EAI enables autonomous environmental interaction and enhances adaptive behavior within aquaculture systems. Following a systematic review methodology (not a meta-analysis), 118 studies from Web of Science, Scopus, IEEE Xplore, PubMed, and Google Scholar were synthesized after screening against predefined eligibility criteria. This synthesis aimed to examine advances across key areas: mobile robotic platforms, sensor fusion, underwater navigation, adaptive control, and species-specific monitoring applications. The review reveals that most current aquaculture systems are still weakly embodied, operating primarily as mobile sensing platforms with onboard or offboard inference rather than truly autonomous learning agents. Despite this, autonomous underwater vehicles (AUVs), remotely operated vehicles (ROVs), uncrewed surface vehicles (USVs), and bioinspired robotic systems significantly enhance spatial monitoring coverage and adaptive environmental sampling compared to stationary sensing methods. However, critical operational barriers persist, including underwater image degradation, biofouling, turbidity, light attenuation, sensor drift, limited battery endurance, and challenges in underwater simultaneous localization and mapping (SLAM) and real-time communication. The review further emphasizes that aquaculture is a highly non-linear biological system where EAI-driven adaptive sampling, multimodal sensor fusion, and reinforcement-learning-based control are particularly effective for improving dissolved oxygen (DO) monitoring, feeding optimization, environmental prediction, and welfare assessment. The study also distinguishes classical feedback control from learning-based embodied agents, offering a critical assessment of the current technological maturity of closed-loop aquaculture automation. Looking ahead, future next-generation aquaculture systems are expected to integrate EAI with Agentic AI, digital twins, energy-aware robotics, and circular-economy principles. This integration aims to reduce feed waste, operational costs, and carbon footprints, thereby improving sustainability and ecological resilience.