Deep Learning for Fruit and Vegetable Detection in Autonomous Harvesting: A Comparative Narrative Review (2014–2026)

Abstract

Autonomous harvesting is central to reducing labour dependence and post-harvest loss in fruit and vegetable production, and fast, accurate visual detection remains its principal bottleneck. This paper systematically reviews deep-learning (DL) approaches to fruit and vegetable detection for robotic harvesting, updating and extending the existing literature in two respects. First, it synthesizes results across 100+ primary studies using a common comparative framework — rather than a chronological narrative — spanning two-stage detectors (R-CNN, Faster R-CNN), one-stage CNN detectors (SSD, YOLOv1–v7), the current anchor-free YOLO family (YOLOv8–YOLOv13), transformer-based real-time detectors (RT-DETR and its derivatives), and instance-segmentation/pose-based approaches used for picking-point localization rather than mere bounding-box detection. Second, it extends the review's scope beyond detection accuracy to the two factors that determine real-world deployability: edge/embedded inference performance and dataset standardization. The review finds that anchor-free YOLO variants (v8–v10) currently offer the best accuracy–latency trade-off for on-robot deployment, that transformer-based detectors (RT-DETR family) match or exceed YOLO accuracy in heavily occluded, cluttered canopy scenes but at a higher compute cost that constrains edge use without quantization or pruning, and that occlusion, illumination variability, and the scarcity of standardized, richly annotated field datasets remain the field's central unsolved problems. We conclude with a structured research agenda covering multimodal (RGB-D/LiDAR) fusion, foundation-model-assisted annotation, cross-domain generalization, and reporting standards that would allow fair comparison across studies — a gap that continues to limit reproducibility in this literature.

Keywords: harvesting robot; fruit and vegetable detection; deep learning; YOLO; RT-DETR; instance segmentation; picking-point localization; edge deployment; precision agriculture