Conference Agenda

Rising global food demand, increasing production costs, and labor shortages present significant challenges for specialty crop production, particularly for labor-intensive tasks such as fruit harvesting. Robotic harvesting offers a promising long-term solution, yet adoption in orchard environments remains limited due to unstructured conditions, variable lighting, and challenges in fruit detection and manipulation. This paper presents an improved autonomous apple harvesting system, Orchard roBot (OrBot), developed by the Robotics Vision Lab at Northwest Nazarene University.

The updated OrBot integrates a dual-camera vision architecture consisting of an eye-to-hand stereo camera with a wide field of view for fruit detection and an eye-in-hand RGB-D camera for precise manipulation. The control system was redesigned using Robot Operating System 2 (ROS2) and Python, enabling modular and coordinated operation. Fruit detection was implemented using a YOLOv5 deep learning model with visual servoing for manipulation guidance.

Laboratory tests achieved a 100% harvesting success rate, while field trials in a commercial Idaho orchard achieved a 75–80% success rate. Results demonstrate improved fruit search efficiency and harvesting reliability while identifying key challenges for future improvement.

This study develops a computer-vision-based automatic grading system for senryo (Sarcandra glabra) to reduce labor requirements and improve grading consistency in agricultural production. Conventional grading relies heavily on manual inspection, resulting in high workload and variability among workers. Based on grading criteria obtained from farmers, the proposed system evalu-ates senryo quality by detecting the number of clusters and the number of seeds within each cluster.
A two-dimensional and three-dimensional framework is being developed. In the 2D stage, a YOLOv5-based detection model was trained to identify seeds and clusters using images collected from online sources, artificial sam-ples, and real flowers purchased directly from farms. To address challenges such as small object size, dense overlap, and class imbalance, data augmenta-tion, small-object-oriented training settings, loss weight adjustment, and hard negative mining were applied. The current model achieved a precision of 0.744, recall of 0.842, and mAP@0.5 of 0.842, demonstrating the feasibility of automated detection for grading support.
In parallel, stereo image acquisition using dual ZED cameras has been initi-ated to enable future 3D-based counting and grading refinement.

Viticulture is one of the first application domains to adopt robotic solutions in agriculture. This sector is particularly promising for the application of autonomous vehicles because of the large number of crop interventions, which require intensive use of manpower, and the relatively high investment capacity of wine-producing companies. In many cases vines are grown on steeply sloped terrain. Operating in such environments requires designing robots capable of safely travel in large transverse sloped terrain. Alpirobot has developed a tracked robot conceived to link different implements and able to automatically adjust its attitude and mass distribution to manage the load repartition between the tracks maximizing the rollover angle.

The proposed solution relies on two independent control systems for adjusting the asset of the main vehicle body: a tilting and a transverse translation mechanism. The tilt system maintains the body close to the vertical alignment, whereas the second system controls the position to optimize the load distribution between the tracks.

This approach reduces soil compaction and rollover risk, thereby enhancing lateral stability and the overall maneuverability of the robotic vehicle.

Vehicle features and preliminary results achieved in lab-controlled and field tests are presented.

Autonomous navigation in orchards requires holistic perception of obstacles and immediate surroundings. Past studies focused mostly on path finding through orchard alleys and obstacle detection using bounding-boxes, limiting explicit distinction between foreground objects, traversable space, and background, necessary for optimal robot navigation. In this work we propose an instance segmentation-based approach for precise delineation of object boundaries in an apple orchard where the latest YOLO series of nano models, YOLO26, YOLO11, and YOLOv8 were applied to compare the performances. A human annotated comprehensive dataset was developed for training, covering multiple orchard object classes including trees, vehicles, alleyways, sky, and irrigation system. Results showed that YOLO11 has better visual understanding compared to YOLOv8 and YOLO26 variations, achieving a mask mAP50 of 0.769 and mAP50-95 of 0.522 with faster inference of 2.06ms, and strong per-class mAP50 scores: Trees (0.96), Sky (0.97), DirtRoad(0.97), and Alley (0.99). YOLO11 slightly outperformed YOLO26 (mAP50-95: 0.50); better capturing small, occluded and occasional object classes. The results show that our approach can be used to effectively distinguish key features across multiple object classes in outdoor orchard environments, enabling reliable detection of safe traversal areas for autonomous vehicles in orchards within tree rows, headlands, and surrounding plots.

Unlike field crops such as wheat and corn with wide row spacing and upright plant type, Hunan mustard, a characteristic leafy crop, faces prominent difficulties: large overlapping leaves, early row closure, narrow row spacing (15–30 cm), strong interference from similar broad-leaved weeds, rapid growth posture changes, and high sensitivity to light and surface environment. These issues make traditional row alignment methods unreliable. To address this, this study proposes a precise mustard row recognition model based on improved YOLOv8n and lightweight semantic segmentation, with subsequent expansion to integrate disease and pest identification as well as growth and development recognition. Innovations include filling the technical gap in dense leafy crop navigation, realizing plant-level semantic recognition without relying on bare inter-rows, and designing a lightweight model for on-board embedded platforms. Field experiments show the model achieves 95.2% recognition accuracy, 30 ms single-frame processing time, and ≤ 3cm lateral deviation, meeting intelligent agricultural machinery requirements and providing technical support.

Accurate estimation of grape composition is essential for site-specific management and optimized harvest planning. This study aims to investigate the use of UAV-based imaging, combined with digital soil mapping and cumulative growing degree days (GDD), within a machine learning to predict grape quality parameters.

The experiment was conducted in a vineyard in southern Italy over two growing seasons. UAV multispectral and thermal imagery was acquired from fruit set to ripeness to characterize canopy status. Climatic data was monitored using an on-site weather station to compute cumulative GDD and model ripening dynamics. Soil variability was characterized using a soil optical sensor to assess organic matter and texture. At harvest, yield components, total soluble solids (TSS), and total acidity were measured.

Multisource UAV, soil, and temperature variables were incorporated into machine learning models. The inclusion of cumulative GDD improved prediction accuracy for both TSS and acidity, while soil data further enhanced model performance. Feature importance analysis indicated that cumulative temperature, terrain slope, and soil properties were among the most relevant factors influencing grape quality variability.

These results demonstrate that combining UAV sensing, temperature-based ripening modelling, and machine learning enables reliable, non-destructive prediction of berry composition to support harvest management in precision viticulture.