Emerging Trends in Agricultural Robotics: Integration of Artificial Intelligence, Machine Vision, and IoT in Farm Machinery

Agricultural robotics is rapidly maturing from isolated prototypes to integrated systems that combine artificial intelligence (AI), machine vision, and the Internet of Things (IoT). This review synthesizes recent advances, practical deployments, and research challenges at the intersection of robotics, perception, and connectivity in farm machinery. We examine design paradigms for autonomous platforms, sensor suites and vision pipelines for crop/weed/fruit tasks, edge and cloud AI for perception and planning, IoT architectures for fleet and energy management, human-robot interaction and safety, and socio-economic aspects including business models and policy. Representative industrial examples and high-impact demonstrations are discussed to highlight real-world readiness and remaining gaps. Finally, we propose a research agenda—covering duty-cycle datasets, standardized benchmarks, robust perception under agricultural conditions, low-power edge AI, and interoperable IoT standards—that aims to accelerate responsible deployment of robotic systems in diverse farm contexts.

Introduction

Automation and robotics in agriculture aim to address pressing challenges: labor shortages, the need for precision to reduce agrochemical use, and sustainability goals requiring higher resource-use efficiency. Over the last decade, improvements in sensing (high-resolution RGB, multispectral, hyperspectral cameras), compute (edge GPUs, NPUs), and AI (deep learning for detection, segmentation, and control) have enabled robots to perform complex field tasks such as targeted weeding, fruit picking, and autonomous tillage. At the same time, IoT connectivity allows distributed fleets and implements to share state, telemetry, and task schedules—shifting systems from single-robot tools to orchestrated, data-driven platforms. Recent reviews summarize the rapid technology diffusion and identify an industry moving from lab prototypes to commercial deployments.

This review focuses on the integration of AI, machine vision, and IoT in farm machinery— how these components are combined, what capabilities they unlock, and what technical and non-technical constraints remain. We consider both stationary (sorting lines, packhouses) and mobile systems (ground robots, UAVs), giving emphasis to mobile field robotics where perception, control, and connectivity interplay strongly. Architectural paradigms for agricultural robotic systems Robotic farm systems broadly implement one of three architectural paradigms: 1. Monolithic autonomous machines: single platforms with onboard sensing, perception, planning, and actuation (e.g., autonomous mowers, laser weeders).

These systems minimize dependency on external connectivity but must carry more compute and energy resources. Examples include the Carbon Robotics Laser Weeder and OEM autonomous tractors. 2. Edge-cloud hybrid systems: perception and low-latency control run on edge compute while higher-level planning, fleet coordination, and model updates are handled in the cloud. This balances real-time responsiveness with centralized optimization and learning.

NAAS Rating (2025): 4.69 IJAN 2025; 7(10): 70-74 Yuvraj Gopinath Kasal Dept. of Farm Power and Machinery, Maharana Pratap Horticultural University, Karnal, Haryana, India Satyapal Singh Dept. of Vegetable Science, Maharana Pratap Horticultural University, Karnal, Haryana, India Shahroon Khan Dept. of Fruit Science, Maharana Pratap Horticultural University, Karnal, Haryana, India Yuvraj Gopinath Kasal Dept. of Farm Power and Machinery, Maharana Pratap Horticultural University, Karnal, Haryana, India https://www.agriculturejournal.net 3. Distributed multi-agent systems: swarms or cooperative fleets (tractor + implements + small robots + drones) that divide labor and communicate via IoT middleware to optimize coverage and energy usage. Cooperative multi-machine operations are attracting research interest and early pilots. Key architectural design decisions depend on task latency, bandwidth availability in rural areas, energy budgets, and safety/regulatory constraints.

Perception: machine vision systems and AI pipelines Sensors and modalities Machine vision in agriculture uses a mix of sensors:  RGB cameras for color and texture cues (fruit detection, weed/plant boundaries).  Multispectral/hyperspectral sensors for physiological and stress detection (NDVI, disease signatures).  Thermal imagers for water stress or animal detection.  LiDAR and stereo vision for 3-D shape, obstacle avoidance and canopy structure estimation.  Time-of-flight/ultrasonic sensors for short-range range finding and collision avoidance. Practical deployments often fuse modalities—e.g., RGB + NIR + depth—to improve robustness under variable illumination and occlusion. Sensor selection balances information needs, weight, power, and cost. AI for detection, segmentation, and pose estimation Deep learning has become the dominant approach for object detection and segmentation in agriculture.

Key capabilities include:  Object detection (Faster-RCNN, YOLO families) for fruits, weeds, pests, and machinery parts.  Instance and semantic segmentation (Mask R-CNN, DeepLab) for precise localization required in harvesting and selective spraying.  Pose estimation and grasp planning for manipulators (keypoint-based or end-to-end grasp networks), crucial for reliable harvesting.  Domain adaptation and few-shot learning to handle varying cultivars, lighting, and growth stages; these reduce the need for prohibitive labeled datasets. Performance remains sensitive to occlusions (leaves obscuring fruit), variable lighting, and high intra-class visual variation, motivating research into robust training, synthetic data augmentation, and active learning approaches. Recent surveys document strong progress but emphasize the gap between lab accuracy and field robustness. Real-time constraints and edge AI Field robots often require real-time perception on power- and space-constrained platforms.

Edge AI strategies include:  Model compression (pruning, quantization) to reduce model size and inference latency.  Efficient architectures (Mobile Net, Efficient Det, Tiny- YOLO) tailored to embedded GPUs/NPUs.  Pipeline optimization—frame skipping, region-of- interest cropping, and sensor-triggered capture—to trade off fidelity for compute savings. Advances in low-power accelerators (embedded GPUs, dedicated NPUs) make higher-accuracy models viable on- field; yet there is a continuous need to balance accuracy, latency, and energy. Control, planning, and manipulation Motion planning and navigation Autonomous navigation combines GNSS (RTK for cm-level accuracy in many precision tasks), visual odometry, LiDAR SLAM, and IMU fusion. Agricultural fields present unique challenges: repetitive textures, featureless bare soil, dynamic obstacles (workers, animals), and challenging GNSS conditions under tree canopies.

Robust systems therefore fuse multiple modalities and incorporate field-aware path planners that respect crop geometry and minimize soil compaction. Manipulation for harvesting and precision tasks Manipulation tasks require gentle, adaptable grippers and compliant control to avoid crop damage. Control approaches include:  Impedance and force control for safe interaction with fruit/plant tissue.  Visual servoing—closing the perception-to-actuation loop with image feedback—to correct for perception and actuation errors.  Learning-based controllers that map vision to end- effector commands (end-to-end or hybrid), particularly for variable fruit poses. Despite progress, harvest robotics still suffer from cycle times and damage rates that lag human pickers in many crops; success has been strongest in high-value, uniform crops (e.g., strawberries under controlled environments) and for mechanical thinning/targeted pruning tasks.

IoT and systems integration: connectivity, data, and fleet management IoT architectures for farm machinery IoT enables telemetry, remote monitoring, OTA model updates, and fleet orchestration. Architectures typically include:  Device layer: sensors and actuators with local controllers and edge compute.  Communication layer: LoRaWAN, NB-IoT, LTE/4G/5G, or Wi-Fi depending on bandwidth and latency needs. Low-bandwidth options suit periodic telemetry; higher bandwidth links or mobile gateways support offloading of rich sensor data (e.g., images) to the cloud.  Cloud services: data lakes, model training, fleet dashboards, and command-and-control APIs.  Orchestration layer: scheduling, task assignment, and energy-aware routing for multiple machines. Challenges include rural connectivity gaps, heterogeneous device ecosystems, and cybersecurity/safety concerns.

Edge-cloud co-design (local autonomy with cloud coordination) mitigates bandwidth limits while enabling centralized optimization. Data pipelines, digital twins and predictive maintenance IoT telemetry enables digital twins—virtual models of machines and fields—that support simulation, what-if https://www.agriculturejournal.net planning, and predictive maintenance (anomaly detection on vibration, motor current, battery metrics). Predictive models reduce downtime and optimize service scheduling, which is especially important where service centers are distant. Standards, interoperability and data governance Interoperability standards (similar to ISOBUS for implements) are emerging as critical for multi-vendor fleets.

Data ownership, privacy, and policies around agricultural data sharing remain active topics—particularly given value in crop yield forecasting, input optimization, and traceability. Recent reviews emphasize the need for open interfaces and clear governance mechanisms to foster trust and wider adoption. Real-world deployments and industrial examples Several high-impact commercial demonstrations in recent years illustrate technological maturity and practical tradeoffs:  Laser Weeder (Carbon Robotics): An AI-driven, GPU-backed tractor-pulled unit that identifies and removes weeds at scale using lasers—demonstrates extreme processing needs and safety concerns associated with high-power actuation in the field. This system highlights the role of cloud/edge GPU farms for model training and high throughput inference.  OEM autonomy and retrofit kits (e.g., Deere): Major manufacturers are introducing autonomous tractors and retrofit autonomy kits that combine multi-camera sensing, RTK GNSS, and centralized fleet management—showing that large-scale industry is embracing autonomy for commercial operations.  Startups & diversified platforms: From small weeding robots to multi-purpose platforms for mowing, spraying, and surveillance, a vibrant startup ecosystem reveals varied business models—robot-as-a-service, implement retrofit, and licensed autonomy stacks.

Industry reports note continued investment growth despite macroeconomic headwinds. These deployments underscore practical constraints—capital expense, safety/regulatory compliance (e.g., lasers, failsafe autonomy levels), and the need for reliable service/support models. Human-robot interaction, safety, and socio-economic considerations Safety and regulatory frameworks Field robots operate near humans and livestock; safety requires redundant sensing, predictable behavior, emergency stop mechanisms, and rigorous validation. Regulatory frameworks for autonomous vehicles have begun to extend into off-road domains, but many jurisdictions lack clear standards for farm robots; this creates uncertainty for commercial adoption.

Labor, skills, and adoption pathways While robotics can alleviate labor shortages, adoption pathways must consider worker reskilling—farm technicians will need skills in robotics maintenance, data management, and safety protocols. Business models such as Custom Hiring Centers (CHCs), robot-as-a-service, or cooperatives can lower farmer capital barriers and improve utilization rates for expensive robotic assets. Ethical and environmental concerns Automation decisions affect rural employment and land management. Careful evaluation of social impact, equitable access to technology, and environmental tradeoffs (e.g., energy use, embodied emissions) must accompany technical development.

Key technical challenges and research opportunities We identify major areas where research would yield high impact: 1. Robust perception under real-world agricultural conditions: long-tail cases (dense occlusion, rain/dust, lighting extremes) remain a major failure mode. Research in sensor fusion, domain adaptation, and self- supervised learning can reduce labeling needs and increase robustness. 2. Standardized datasets and benchmarks: the community needs representative, annotated datasets covering multiple crops, growth stages, and environmental conditions to benchmark detection, segmentation, and manipulation.

Public benchmarks will drive reproducibility and accelerate progress. 3. Low-power, low-latency edge AI: optimized model architectures, compiler toolchains, and hardware- software co-design for embedded platforms that balance inference accuracy with energy and thermal budgets. 4. Multi-agent coordination and task allocation: scalable algorithms for heterogeneous fleets that can share perception, coordinate coverage, and dynamically reassign tasks considering energy and operational constraints. 5. Interoperable IoT and data governance: open APIs and standardized telemetry schemas to enable third- party tools, avoid vendor lock-in, and enable secure, privacy-preserving data sharing. 6.

Field-ready manipulation and soft robotics: compliant, adaptive end-effectors and tactile sensing for delicate harvesting tasks, with control strategies that generalize across fruit geometries and attachment modes. 7. Validation frameworks and safety cases: standardized testbeds and simulation-to-real validation pipelines for proving safety in mixed human-robot workspaces. Roadmap and recommendations For researchers, industry, and policymakers we recommend:  Short term (1-2 years): Establish open datasets and shared evaluation protocols; pilot edge-cloud setups in cooperative farms; fund workforce training programs for robot maintenance.  Medium term (3-5 years): Develop fleet orchestration tools and interoperable IoT standards; deploy predictive maintenance and digital-twin pilots in commercial settings; support regulatory clarity around safety and certification.  Long term (>5 years): Mature modular robotic platforms with swappable payloads; mainstream robotic services via CHCs and coop models; integrate robotics outcomes into sustainability policy (reduced agrochemical use, improved yields, traceability). https://www.agriculturejournal.net Table 1: Comparison of sensing modalities for common agricultural perception tasks Sensing Modality Strengths Limitations Typical Cost (USD) RGB Cameras 100-500 Low cost, easy integration, high spatial resolution; suitable for fruit detection and weed mapping Sensitive to lighting variation, limited spectral information, poor performance under shadows Multispectral Cameras 1,000-5,000 Enables vegetation index computation (NDVI, GNDVI); good for disease detection and crop health monitoring Moderate cost; limited spectral range (typically 4-6 bands); sensitive to calibration errors Hyperspectral Sensors High spectral resolution enables early disease detection, nutrient stress monitoring Expensive, data-intensive, requires expert calibration and analysis 10,000- 50,000 Thermal Infrared Cameras Useful for irrigation monitoring, canopy temperature mapping, stress detection Lower spatial resolution; influenced by ambient temperature; requires calibration 2,000-10,000 LiDAR (Light Detection and Ranging) Accurate 3D structure and canopy mapping; robust in low light High cost, heavy payload, data processing intensive 5,000-25,000 Time-of-Flight (ToF) Sensors / Depth Cameras Low-cost 3D sensing, effective for fruit localization and robot navigation Short range (<5 m); affected by sunlight and reflective surfaces 300-2,000 Ultrasonic Sensors Simple, robust for obstacle detection, canopy height estimation Low resolution, wide beam divergence, not suitable for fine object detection 50-300 Radar Sensors (mmWave, UWB) Works in all weather/light conditions, good for obstacle avoidance and terrain sensing Coarse resolution, requires complex signal processing 200-2,000 Spectroradiometers Precise spectral data for biochemical analysis, disease stress quantification Very high cost, unsuitable for mobile field robots 15,000- 60,000 Combination (Sensor Fusion) Integrates RGB + LiDAR + multispectral for enhanced robustness and accuracy Increased complexity, synchronization challenges, higher total cost 5,000- 50,000+ Table 2: Representative robotic applications, maturity level, key enabling technologies, and adoption challenges Robotic System/ Application Primary Function Maturity Level Key Enabling Technologies Major Adoption Challenges LaserWeeding (Carbon Robotics, USA) Precision weed removal using laser targeting Commercial Deep learning-based weed recognition, high-power CO₂ laser actuation, autonomous navigation Safety regulations for lasers, high capital cost, field safety certification See & Spray (John Deere) Targeted herbicide application Commercial Machine vision, CNN-based plant classification, real-time actuation control Integration with existing sprayers, lighting variation, data labeling requirements Naïo Oz and Dino (Naïo Technologies, France) Autonomous mechanical weeding and hoeing Commercial RTK-GPS navigation, machine vision, IoT-based fleet monitoring High purchase cost, limited crop adaptability, maintenance training Thorvald II (University of Lincoln, UK) Prototype / Early Commercial ROS-based control, LiDAR + RGB fusion, modular chassis Modular multi-purpose robot (weeding, spraying, phenotyping) Swarm-based seeding and crop monitoring Prototype / Field Testing IoT communication, GNSS guidance, distributed AI control Lack of standardization, interoperability, limited local support Fendt Xaver Swarm Robots (AGCO, Germany) Data synchronization, scalability, connectivity in rural areas Robocrop Vision Guidance (Garford, UK) Vision-based precision hoeing Commercial High-speed image processing, color segmentation, adaptive control Limited to row crops, sensitivity to canopy occlusion Ecorobotix ARA Smart spot-spraying and weed targeting Commercial AI-based image segmentation, GPS guidance, low-volume precision dosing Research / Prototype SLAM navigation, multi-sensor fusion, AI-based decision support Cost vs. savings balance, weed misclassification in dense canopy AgBot II (Queensland University of Technology, Australia) Autonomous scouting, soil sampling, and weeding Ruggedization, regulatory approval for autonomous mobility Harvest CROO Robotics (USA) Automated strawberry harvesting Prototype / Pilot Commercial 3D vision, robotic grasping, soft robotics actuators Handling delicacy, speed, variability in fruit position Octinion Rubion Robotic strawberry picker with soft gripper Commercial (Limited) 3D stereo vision, deep learning fruit detection, pneumatic soft actuation Throughput limitations, mechanical complexity, cost recovery Commercial AgXeed AgBot (Netherlands) Legal framework for driverless operation, high upfront cost Autonomous tractor platform for field operations AI-driven route planning, LiDAR safety systems, cloud-based telematics FarmDroid FD20 (Denmark) Solar-powered seeding and weeding robot Commercial GNSS-based navigation, energy- efficient embedded control Dependence on solar energy, limited payload capacity Blue River LettuceBot Selective thinning and spraying of lettuce Commercial (John Deere) High-speed image classification, precision nozzle control Cost of maintenance, adaptation to other crops SwagBot (Australia) Prototype Vision + LiDAR navigation, IoT- based animal tracking Pasture monitoring and autonomous livestock management Battery endurance, terrain adaptability, connectivity constraints https://www.agriculturejournal.net

Robotic System/ Application	Primary Function	Maturity Level	Key Enabling Technologies	Major Adoption Challenges
LaserWeeding (Carbon Robotics, USA)	Precision weed removal using laser targeting	Commercial	Deep learning-based weed recognition, high-power CO₂ laser actuation, autonomous navigation	Safety regulations for lasers, high capital cost, field safety certification
See & Spray (John Deere)	Targeted herbicide application	Commercial	Machine vision, CNN-based plant classification, real-time actuation control	Integration with existing sprayers, lighting variation, data labeling requirements
Naïo Oz and Dino (Naïo Technologies, France)	Autonomous mechanical weeding and hoeing	Commercial	RTK-GPS navigation, machine vision, IoT-based fleet monitoring	High purchase cost, limited crop adaptability, maintenance training
Thorvald II (University of Lincoln, UK)	Modular multi-purpose robot (weeding, spraying, phenotyping)	Prototype / Early Commercial	ROS-based control, LiDAR + RGB fusion, modular chassis	Lack of standardization, interoperability, limited local support
Fendt Xaver Swarm Robots (AGCO, Germany)	Swarm-based seeding and crop monitoring	Prototype / Field Testing	IoT communication, GNSS guidance, distributed AI control	Data synchronization, scalability, connectivity in rural areas
Robocrop Vision Guidance (Garford, UK)	Vision-based precision hoeing	Commercial	High-speed image processing, color segmentation, adaptive control	Limited to row crops, sensitivity to canopy occlusion
Ecorobotix ARA	Smart spot-spraying and weed targeting	Commercial	AI-based image segmentation, GPS guidance, low-volume precision dosing	Cost vs. savings balance, weed misclassification in dense canopy
AgBot II (Queensland University of Technology, Australia)	Autonomous scouting, soil sampling, and weeding	Research / Prototype	SLAM navigation, multi-sensor fusion, AI-based decision support	Ruggedization, regulatory approval for autonomous mobility
Harvest CROO Robotics (USA)	Automated strawberry harvesting	Prototype / Pilot Commercial	3D vision, robotic grasping, soft robotics actuators	Handling delicacy, speed, variability in fruit position
Octinion Rubion	Robotic strawberry picker with soft gripper	Commercial (Limited)	3D stereo vision, deep learning fruit detection, pneumatic soft actuation	Throughput limitations, mechanical complexity, cost recovery
AgXeed AgBot (Netherlands)	Autonomous tractor platform for field operations	Commercial	AI-driven route planning, LiDAR safety systems, cloud-based telematics	Legal framework for driverless operation, high upfront cost
FarmDroid FD20 (Denmark)	Solar-powered seeding and weeding robot	Commercial	GNSS-based navigation, energy- efficient embedded control	Dependence on solar energy, limited payload capacity
Blue River LettuceBot	Selective thinning and spraying of lettuce	Commercial (John Deere)	High-speed image classification, precision nozzle control	Cost of maintenance, adaptation to other crops
SwagBot (Australia)	Pasture monitoring and autonomous livestock management	Prototype	Vision + LiDAR navigation, IoT- based animal tracking	Battery endurance, terrain adaptability, connectivity constraints

Table text extracted from the original PDF; superscript letters are shown inline. The original table image is below.

View original table image

↗ Enlarge

Table 2Close ×

Representative robotic applications, maturity level, key enabling technologies, and adoption challenges

Open full-size image ↗

Original paper tableSource PDF · page 4 ↗

Sensing Modality	Strengths	Limitations	Typical Cost (USD)
RGB Cameras	Low cost, easy integration, high spatial resolution; suitable for fruit detection and weed mapping	Sensitive to lighting variation, limited spectral information, poor performance under shadows	100-500
Multispectral Cameras	Enables vegetation index computation (NDVI, GNDVI); good for disease detection and crop health monitoring	Moderate cost; limited spectral range (typically 4-6 bands); sensitive to calibration errors	1,000-5,000
Hyperspectral Sensors	High spectral resolution enables early disease detection, nutrient stress monitoring	Expensive, data-intensive, requires expert calibration and analysis	10,000- 50,000
Thermal Infrared Cameras	Useful for irrigation monitoring, canopy temperature mapping, stress detection	Lower spatial resolution; influenced by ambient temperature; requires calibration	2,000-10,000
LiDAR (Light Detection and Ranging)	Accurate 3D structure and canopy mapping; robust in low light	High cost, heavy payload, data processing intensive	5,000-25,000
Time-of-Flight (ToF) Sensors / Depth Cameras	Low-cost 3D sensing, effective for fruit localization and robot navigation	Short range (<5 m); affected by sunlight and reflective surfaces	300-2,000
Ultrasonic Sensors	Simple, robust for obstacle detection, canopy height estimation	Low resolution, wide beam divergence, not suitable for fine object detection	50-300
Radar Sensors (mmWave, UWB)	Works in all weather/light conditions, good for obstacle avoidance and terrain sensing	Coarse resolution, requires complex signal processing	200-2,000
Spectroradiometers	Precise spectral data for biochemical analysis, disease stress quantification	Very high cost, unsuitable for mobile field robots	15,000- 60,000
Combination (Sensor Fusion)	Integrates RGB + LiDAR + multispectral for enhanced robustness and accuracy	Increased complexity, synchronization challenges, higher total cost	5,000- 50,000+

Table text extracted from the original PDF; superscript letters are shown inline. The original table image is below.

View original table image

↗ Enlarge

Table 1Close ×

Comparison of sensing modalities for common agricultural perception tasks

Open full-size image ↗

Original paper tableSource PDF · page 4 ↗

Conclusions

Integration of AI, machine vision, and IoT is propelling agricultural robotics from demonstrations to real impact in the field. Technical advances across perception, planning, manipulation, and systems integration have enabled novel capabilities—precision weeding, autonomous harvesting, and fleet orchestration—while also exposing challenges in robustness, energy, connectivity, and safety. Addressing these challenges will require multidisciplinary collaboration across robotics, agronomy, data governance, and policy. With careful attention to social and environmental tradeoffs and by building interoperable, resilient systems, agricultural.

Acknowledgments

The author thanks colleagues at the College of Horticulture, M.P.H.U., Karnal, and the broader agricultural robotics community for discussions and inputs.

1. Spagnuolo M, et al. Agricultural robotics: A technical review addressing applications and market trends. Robotics (MDPI). 2025;14(2):9.
2. Thakur A, et al. An extensive review on agricultural robots with focus on machine vision techniques. Comput Electron Agric. 2023.
3. Sharma K, et al. Integrating artificial intelligence and Internet of Things in modern agriculture. Elsevier Rev. 2024.
4. Miller T, et al. The IoT and AI in agriculture: The time is now. Front Syst (Open Access Rev). 2025.
5. Carbon Robotics (LaserWeeder) demonstration coverage. Tom’s Hardware; 2025.
6. Reuters. Deere boosts bet on autonomous tractors with new machines. Reuters News Report. 2025.
7. Prabin Kumar Sahoo, Deoraj Sharma. Economic impact of artificial intelligence in the field of agriculture. Int J Hortic Food Sci 2023;5(1):29-34. DOI: 10.33545/26631067.2023.v5.i1a.152
8. Kumar V, et al. A comprehensive review on smart and sustainable IoT for agriculture; 2024.
9. Kasal YG, Singh S, Khan S. Recent trends in farm machineries for horticulture. Int J Agric Food Sci. 2025;7(9):909- 11. https://doi.org/10.33545/2664844X.2025.v7.i9l.843
10. Liu L, et al. Smart sensors and IoT solutions for precision agriculture; 2025.