Session 5B – Agricultural and Field Robotics
Date: Thursday, 19th November 2026
Time: 14:00 – 15:30
Room: Aula Capella
Session Chair: TBA
Presentation format: Each paper is allocated 15 minutes: approximately 10–12 minutes for the presentation and 3–5 minutes for questions.
14:00 – Optimization-based Multi-Robot Task Allocation for Human-Robot Collaboration in Vineyard Harvesting
Authors: Dimitrios Troullinos¹, Maria Nuria Conejero², Filippo Bistaffa¹, Jose M. Bengochea-Guevara³, Angela Ribeiro³ and Juan A. Rodriguez-Aguilar¹
Affiliations: ¹ Artificial Intelligence Research Institute (IIIA-CSIC), Barcelona, Spain; ² Escuela Técnica Superior de Ingenieros Industriales (UPM), Madrid, Spain; ³ Center for Automation and Robotics (CAR-UPM-CSIC), Madrid, Spain.
Abstract
Agricultural robotics is increasingly relevant across a wide range of applications. Especially in Human-Robot Collaboration settings, there are many opportunities to leverage existing technological infrastructure and enable robots to assist humans with labor-intensive tasks. Motivated by an existing mobile robotic platform for vineyard harvesting, in which wheeled robots follow human operators and transport grapes, we focus on coordinating the robotic fleet so that operators are not left waiting for an available robot. We frame this as a Multi-Robot Task Allocation problem and formulate it as a Mixed-Integer Linear Programming model that schedules and routes the robots through the vineyard, minimizing idle times of human operators and respecting robots’ exclusive use of narrow yard corridors. The experimental evaluation in simulation demonstrates the fleet’s coordination capabilities. Furthermore, we examine a fleet-configuration problem that automatically adjusts the robotic fleet’s size to meet a given quality-of-service requirement with the minimum number of robots.
14:15 – VineComputer: Automated Bud Forcing for Early Yield Potential Assessment
Authors: Miguel Oliveira¹, Germano Moreira², Filipe Neves dos Santos¹, Mário Cunha² and André Baltazar³
Affiliations: ¹ INESC TEC and Faculty of Engineering, University of Porto, Porto, Portugal; ² INESC TEC and Faculty of Sciences, University of Porto, Porto, Portugal; ³ INESC TEC, Porto, Portugal.
Abstract
The growing demand for efficient and sustainable agriculture, alongside labour shortages and climate variability, has accelerated the adoption of precision viticulture technologies. Early yield assessment is vital for informed viticultural decision-making. This paper presents VineComputer, an automated greenhouse designed to force the budbreak of vine dormant buds, thereby assessing bud fruitfulness and ultimately yield potential well in advance. By automating this traditionally manual procedure, the system is designed to reduce reliance on specialised labour and minimise the high variability inherent in manual counting techniques, addressing critical challenges within the economically significant Portuguese wine sector. The proposed methodology utilises Controlled Environment Agriculture (CEA) to precisely regulate the microclimate and irrigation conditions necessary for budbreak. Furthermore, it integrates automated image acquisition and machine learning techniques to autonomously monitor the tray and classify the state of each cutting (empty, dormant or sprouted). By embedding the bud-forcing technique within a highly controlled and reproducible environment, this automated approach significantly increases both efficiency and reproducibility. Ultimately, this proof of concept shows the feasibility of combining environmental automation with computer vision, marking a meaningful step toward a more resilient, digitised, and technologically advanced viticulture.
14:30 – GeoZigzag Studio: A Reproducible Workflow for Agricultural Coverage and Semantic Georouting
Authors: Luis Prieto-Lopez, Sergio Sanchez de La Fuente, Miguel A. Gonzalez-Santamarta, Angel Manuel Guerrero-Higueras, Lidia Sanchez-Gonzalez and Francisco J. Rodriguez-Lera
Affiliations: University of Leon, Leon, Spain.
Abstract
Agricultural robotic missions combine two complementary planning tasks: field coverage and navigation between geographically distributed targets. Existing workflows typically rely on multiple independent tools, complicating mission preparation and deployment. This paper presents GeoZigzag Studio, a unified framework integrating coverage planning, semantic routing, mission validation, and ROS-oriented mission export. The contribution is an integration and reproducibility layer rather than a new coverage or routing algorithm. In the evaluated scenarios, row spacing and waypoint spacing independently control coverage geometry and waypoint density, semantic routing avoids all declared forbidden regions with a 4.6–7.3% distance increase, and network-routing snap distance should be explicitly considered for agricultural missions. Simulation and preliminary outdoor experiments demonstrate that the generated missions can be consumed by ROS 2 robotic platforms after the required map-frame transformation.
14:45 – End-to-end Monocular Visual-Inertial Odometry For Agricultural Environments
Authors: Víctor Romero Bautista¹, Nelson Broyer Paco Chipana¹, Luis V. Calderita Estévez¹, Leopoldo Altamirano Robles² and Pedro M. Núñez Trujillo¹
Affiliations: ¹ Robolab-Robotics and Artificial Vision, University of Extremadura, Cáceres, Spain; ² Computer Vision Laboratory, INAOE, Puebla, México.
Abstract
Agricultural environments significantly reduce the accuracy of monocular visual-inertial odometry systems compared to indoor and urban environments. This is due to unstructured scenes with unstable features, repetitive patterns, uneven terrain, and drastic changes in lighting. Traditional systems based on manual feature extraction present difficulties in tracking visual features in these types of environments; furthermore, traditional fusing methods raise challenges related to calibration, synchronization, and the fusion of camera and IMU data. Deep learning-based visual-inertial odometry methods offer a way to incorporate visual-inertial information through automatic pattern correlation learned during the training process. In this work, we propose to develop monocular visual-inertial odometry method that address pose estimation problems in unstructured agricultural environments. To this end, we employ deep learning models that extract inertial and spatiotemporal features; these features are then fused using a cross-attention mechanism. Experimental results show that the proposed method performs competitively compared to state-of-the-art monocular visual and monocular visual-inertial odometry methods.
15:00 – LiDAR-Only Dual-ROI Row Perception and Trajectory Generation for Autonomous Inter-Row Vineyard Mowing
Authors: David Salazar Brás¹, Pedro O. Rodrigues², André Silva Aguiar³, Filipe Neves dos Santos¹ and André Rodrigues Baltazar³
Affiliations: ¹ Faculty of Engineering, University of Porto (FEUP), and INESC TEC, Porto, Portugal; ² University of Trás-os-Montes and Alto Douro (UTAD) and INESC TEC, Vila Real / Porto, Portugal; ³ INESC TEC, Porto, Portugal.
Abstract
Mowing vineyard inter-row vegetation is recurrent, labour-intensive, and well suited to autonomous execution. This paper presents a LiDAR-only perception and local-planning stack that keeps a mobile robot at a fixed lateral distance from a vine row without row maps, satellite row positioning, or added field infrastructure. Each corridor side is modelled by two regions of interest at distinct longitudinal horizons, whose oriented crop boxes are re-placed on every sweep from the previously accepted line fit. This predictive placement forms a per-region tracking loop that follows curved rows without an explicit curvature model or dense state estimator. Row lines are extracted by vertical-plane RANSAC with physically interpretable acceptance gates, smoothed by an inlier-ratio-weighted finite impulse response filter with circular averaging, and condensed into a bounded per-side safety score. The planner converts the row model into a continuous path of chained cubic Bézier curves. In simulated corridors with analytic ground truth, the row-line estimate achieves lateral RMSE below 0.10 m in nominal geometries and the planned reference agrees with the ideal path within 0.10 m RMSE; on real-vineyard data, perception runs in under 23 ms per sweep.
15:15 – Semantic Digital Twin Generation of Arbutus Unedo Fruit Trees Using RGB-D Sensing, YOLOv8, and 3D Gaussian Splatting
Authors: Beatriz Dias¹, Sandro Costa Magalhães² and Filipe Neves dos Santos³
Affiliations: ¹ ISEP, UTP, Porto, Portugal; ² INESC TEC and ISEP, UTP, Porto, Portugal; ³ FEUP and INESC TEC, Porto, Portugal.
Abstract
Robotic harvesting requires accurate semantic 3D representations of fruit trees, but orchard environments remain challenging due to occlusions, dense foliage, irregular geometry, variable illumination, and sensor noise. This paper proposes a RGB-D perception pipeline for generating object-centred Digital Twins of fruit trees by integrating OAK-D data acquisition, Structure-from-Motion (SfM), neural scene reconstruction, YOLOv8-based semantic masking, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) fruit localisation, and colour-based maturity classification. The approach was evaluated on Arbutus Unedo fruit trees using 1,835 synchronised RGB-D image pairs and 3D-printed fruits representing different maturity stages. Results show that semantic masking substantially improves object-centred 3D Gaussian Splatting (3DGS) reconstruction, increasing Peak Signal-to-Noise Ratio (PSNR) from 6.66 to 35.48, Structural Similarity Index Measure (SSIM) from 0.005 to 0.994, and reducing Learned Perceptual Image Patch Similarity (LPIPS) from 0.797 to 0.029. In contrast, Neural Radiance Fields (NeRF) showed only marginal improvement under the same semantic guidance. The proposed framework localised nine fruits in 3D and classified them as READY or UNRIPE, demonstrating its potential for semantic Digital Twin generation, selective harvesting, and precision orchard monitoring.