Last month, the BARD team visited Ben-Gurion University of the Negev. We met with Prof. Yael Edan (Department of Industrial Engineering), Prof. Amir Sagi and Prof. Eliahu Aflalo (Department of Life Sciences), who together with Prof. David Zarrouk (Department of Mechanical Engineering) are implementing precision robotic methods to better understand and utilize data for aquacultural purposes.
This project develops collaborative robotic systems to improve sustainability, productivity, and biosecurity across diverse aquaculture environments. As the world’s fastest-growing protein sector, aquaculture faces mounting challenges in both the U.S. and Israel, particularly in monitoring, feeding, and water quality management.
Prof. David Zarouk in the Bioinspired and Medical Robotics lab. https://designandrobotics.weebly.com/
Equipped with advanced sensing, communication, and control algorithms, these robotic platforms enable precise spatial and temporal monitoring of fish growth , and water quality, as well as targeted execution of feeding and husbandry practices. The research is conducted using swim-bots (SAWbots) and autonomous surface vessels (ASVs) designed to operate across larval, nursery, and grow-out phases in indoor, extensive, and coastal facilities.
BARD team observing the SAWbot developed as part of this project
The algorithms for fish and prawn biomass monitoring were developed, supported by extensive data collection during 2025 from seven aquaculture sites in Israel and the U.S. using diverse camera configurations. These datasets enabled robust fish-detection pipelines based on a YOLOv8s deep learning model, with transfer-learning experiments designed to simulate real-world adaptation to new species and environments under limited labeled data. Results showed that pretraining large generic datasets (COCO) is essential, significantly outperforming training from scratch, while sequential fine-tuning from tilapia to striped bass highlighted the challenge of catastrophic forgetting. Freezing model backbones mitigated memory loss but limited adaptability, leading to the conclusion that large, diverse pretraining datasets are preferable to small domain-specific ones. Ongoing research focuses on improving performance with small datasets and reducing the main bottleneck manual data labeling through automated labeling techniques.
For crustaceans, transfer learning proved effective for morphometric estimation, particularly total length (TL), using YOLOv11-Pose to predict carapace and total length of Macrobrachium rosenbergii from keypoints validated against manual measurements from 2024–2025 datasets. Two-stage fine-tuning significantly improved TL accuracy across seasons, while CL showed limited gains, suggesting a performance ceiling. Visibility and turbidity experiments in outdoor ponds and natural waters revealed severe limitations for optical sensing in Israeli ponds compared to U.S. sites, reinforcing the need for complementary sensing modalities such as sonar and infrared imaging. These sensing and learning systems were integrated with water-quality monitoring using multi-parameter sensors (e.g., DO, pH, conductivity, temperature) and evaluated in controlled RAS facilities at BGU and in diverse U.S. environments, across multiple species and water conditions, using both remotely operated and autonomous robotic modes.
This BARD-funded project is conducted in collaboration with Prof. Steven Hall, Director of the Marine Aquaculture Research Center at North Carolina State University, and his team.
