Development of a low-cost ground robot for real-time ArUco marker localization using lightweight deep learning models on an edge device

• Developed a low-cost ground robot to encourage adoption by smallholder farmers • Explored lightweight YOLO models for ArUco marker-based in-row navigation • Detailed and extensive validation of lightweight YOLO models on marker detection • Validated the feasibility of in-row marker-based navigation using field experiments • Benchmarked custom-trained YOLOv8 and YOLOv11 models against OpenCV’s ArUco module on a Raspberry Pi 4 The high cost of autonomous navigation systems in agricultural robots remains a key barrier to the adoption of field robotics among smallholder farmers. To address this, we present the development of the Reduction-to-Below-Two-Grand (R2B2) robot and the use of lightweight models for fiducial marker recognition to achieve in-row robot localization. As a step towards achieving GPS-independent in-row navigation using ArUco markers, two lightweight object-detection networks (YOLOv8n and YOLOv11n) were trained to classify four ArUco markers under varying conditions. A dataset of 2343 images was utilized for training and internal validation, where both models achieved strong performances (mAP50 ≈ 98% and mAP50-95 ≈ 93%). However, YOLOv8n outperformed YOLOv11n in all performance metrics except precision. Furthermore, both YOLO models and OpenCV’s ArUco module were evaluated on an external dataset comprised of 24 indoor and 16 outdoor treatments captured under varying speed of the R2B2-robot, terrain, and lighting conditions. OpenCV’s ArUco module achieved perfect precision and recall in most treatments, while YOLOv8n edged over YOLOv11n and sustained precision ≥ 85% across all outdoor scenarios. However, both models maintained near-perfect recall in both indoor and outdoor treatments, confirming reliable marker detection even at robot’s speed of 1.5 m/s. Benchmarking inference tests for a duration of 20 seconds at various frames per second on Raspberry Pi 4 revealed that OpenCV’s ArUco module offered significantly faster inference with lower computational demand but was less robust to partial occlusion compared to the YOLO models. Between the YOLO models, YOLOv11n achieved faster inference (∼1380 ms/frame) but used more memory (∼427 MiB), while YOLOv8n was slightly slower (∼1500 ms/frame) yet more memory-efficient (∼423 MiB). These results demonstrate that low-cost hardware combined with lightweight deep learning models can provide the accuracy and robustness required for reliable ArUco marker–based navigation in outdoor agricultural environments.