Robotics in Precision Agriculture: Task-, Platform-, and Evaluation-Oriented Review

Abstract

Robotics is increasingly positioned as an enabling technology for precision agriculture, where management actions must be spatially and temporally targeted under constraints on labour, input use, safety, and environmental impact. This review synthesises studies on agricultural field robotics and organises the literature along four complementary axes: task (monitoring, weeding, spraying, and harvesting), platform (UGV, UAV, gantry/fixed-structure, greenhouse robot, and hybrid systems), autonomy-stack module (perception, localisation, planning, control, actuation, safety, and human–robot interaction), and evaluation setting (lab, greenhouse, open-field single season, and open-field multi-season/multi-site). Across these dimensions, this review analyses how platform constraints shape sensing geometry, actuation capability, localisation reliability, energy/endurance, supervision burden, and safety requirements. It further examines enabling technologies that recur across tasks, including vision and multimodal perception under occlusion and illumination variability, localisation and mapping under weak or denied GNSS, uncertainty-aware planning in deformable and partially observed environments, and compliant end-effectors for contact-rich operations. Beyond cataloguing systems, this paper emphasises evaluation practice by synthesising core task-relevant metrics, comparing laboratory and field validation settings, and proposing a reporting checklist and benchmark ladder to improve reproducibility and cross-study comparability. This review identifies recurring bottlenecks in domain shift, long-term autonomy, calibration robustness, crop-safe actuation, and safety assurance near humans, and it concludes with a staged research roadmap linking near-term evaluation reform to longer-term credible multi-site autonomy. Overall, this paper provides a structured framework for interpreting agricultural robotic systems not only by application but also by deployment context, system maturity, and evaluation credibility.

1. Introduction

Meeting the increasing food demand of the world population in the coming decades is one of the central challenges facing agriculture. Recent synthesis studies indicate that total global food demand is expected to rise substantially by 2050, although the magnitude of this increase depends on assumptions regarding population growth, income, diets, and climate. A meta-analysis by van Dijk et al. [1] projected that the total global food demand would increase by 35% to 56% between 2010 and 2050 across five representative socio-economic scenarios, with the range shifting to 30% to 62% when climate change is incorporated. In a later reassessment, Falcon et al. [2] projected a 50% to 60% increase in total global food demand between 2019 and 2050, with a preferred estimate of about 57% under assumptions on population and income growth, which also link rising demand to a global population approaching about 9.7 to 9.8 billion by 2050.

Meeting this growing demand cannot rely on production expansion alone, because agricultural intensification must also be reconciled with environmental sustainability, resource limitations, and climate variability [3,4]. In this context, precision agriculture has emerged as a core Agriculture 4.0 approach based on observing, measuring, and responding to temporal and spatial variability rather than managing entire fields using average conditions [5]. As emphasized by Reina [5], improving crop production under future food demand will require the adoption of new technologies and artificial intelligence to support more sustainable, site-specific, and timely agronomic decisions. Such an approach is intended to improve production efficiency while reducing unnecessary use of labour, water, fertilizers, and agrochemicals.

Precision agriculture refers to the integrated use of hardware and software technologies that enable farmers to make informed and differentiated decisions for agricultural operations, including planting, fertilization, pest control, and harvesting [6]. It relies on timely, spatially referenced measurements and targeted interventions, typically implemented under practical constraints such as labour availability, input costs, and environmental externalities. Robotics is increasingly positioned as an enabling technology for precision agriculture because many agricultural operations must be executed with spatial and temporal selectivity under constraints on labour availability, input costs, and environmental impact [5]. At the same time, agricultural robotic systems operate in environments that are only partially structured, highly variable, and often degraded by occlusion, illumination changes, weak or absent global navigation satellite system (GNSS) signals, and terrain irregularity. Reviews published in recent years therefore show that agricultural robotics cannot be understood only through algorithmic performance but must also be interpreted through platform constraints, sensing conditions, and operational context [7].

Field robotics is studied in this context because robots can couple repeated sensing to localised intervention at scales ranging from plants to rows, including weed control, variable-rate application, and harvesting operations. This is particularly important because the same autonomy module may behave very differently when deployed on unmanned aerial vehicles (UAVs), unmanned ground vehicles (UGVs), gantry and fixed-structure systems, or greenhouse robots. Recent field-robot surveys and systematic reviews report that agricultural robots are being investigated across monitoring, scouting, weed control, plant protection, and harvesting, reflecting both practical demand and the maturity of sensing and mechatronics needed for these tasks [8,9]. Moreover, studies on precision agriculture information gathering in open-field settings using unmanned ground vehicles (UGVs) and unmanned aerial vehicles (UAVs) show that autonomous multi-robot systems are promising for fine spatiotemporal data collection, but practical deployment must balance information gain with energy consumption and data security constraints [6].

Field conditions differ from industrial environments in ways that change both autonomy design and the meaning of reported performance. In open fields and orchards, the task-relevant objects—plants and produce—are deformable, seasonally changing, and frequently self-occluding; illumination varies with time of day and weather; the ground surface introduces slip, sinkage, and vibration that affect both control and sensor stability. Phenotyping-system reviews note that crop growth stages and canopy closure can constrain robot traversal and sensing geometry and that lodging can force intervention strategies or manual operation even when nominal autonomy succeeds [10]. Zapotezny-Anderson et al. [11] identified occlusion handling (e.g., leaves or branches concealing targets) as a common and non-trivial challenge in agricultural robotics, largely because farming environments are unstructured and cluttered compared with typical industrial robot settings. Similarly, reviews such as [12,13,14] identified perception, sensing, and control limitations as major constraints on agricultural robot performance, including fruit localisation errors caused by detection failures, path planning inaccuracies, and unresolved manipulator kinematics, as well as low deep learning inference speeds that limit real-time harvesting. They also reported perception failure modes characterised by the absence of detectable features (e.g., sensor failure or field departure) and steering failures arising from errors in steering command computation, steering actuation, or both while highlighting the need for improved materials and methods for universal modular sensors, along with dynamic recognition and rapid decision-making in complex natural environments.

This review focuses on four tasks, monitoring, weeding, spraying, and harvesting, because they jointly cover the core “sense–decide–act” pathways of precision agriculture. Monitoring encompasses scouting and phenotyping, where the output is a spatially referenced estimate of crop state or anomalies. Weeding and spraying represent selective intervention modalities, differing primarily in actuation (i.e., physical removal or suppression versus controlled deposition of plant protection products). Harvesting represents contact-rich, quality-sensitive manipulation in which perception and planning must culminate in a physical interaction that preserves produce integrity. This task set is consistent with task taxonomies in recent agricultural robotics reviews and enables cross-task synthesis of shared autonomy-stack requirements [7,15].

Platform-based classification is used because platform constraints shape feasible sensing, actuation, autonomy, and evaluation. UGVs can carry tools and payloads to support physical intervention but must navigate row corridors and manage soil–terrain interaction. UAVs provide rapid coverage and a canopy-level viewpoint that supports broad monitoring, but they face endurance limits and regulatory constraints that bound operational envelopes and influence mapping resolution. Gantry and fixed-structure systems trade mobility for repeatability and controlled sensor pose, which can simplify calibration and enable dense temporal sampling, while greenhouse robots exploit partial structure but often operate under GNSS denial and confined-space safety constraints. These platform differences influence both autonomy-stack design and the external validity of evaluations across crops and settings [16,17]. Unmanned aerial systems (UASs) are included explicitly because they are not peripheral to precision agriculture but central to field-scale robotic monitoring and increasingly relevant to selective intervention workflows. Beyond direct sensing roles, UASs materially influence the design of heterogeneous robotic systems by providing georeferenced observations, prescriptions, or priority maps that guide subsequent UGV operations. Their inclusion is therefore necessary not only to represent a major agricultural robotic platform but also to capture the monitor-then-act logic that increasingly underpins UAV–UGV coordination in agricultural environments.

Across tasks and platforms, recurring bottlenecks appear across the autonomy stack. Perception robustness is frequently limited by domain shift across season, cultivar, illumination, and acquisition protocol; robustness-focused crop imaging reviews argue that credible claims require evaluation designs that hold out sites and seasons rather than relying on random splits [18,19]. Localisation can degrade under canopy, dust, and repeated textures, motivating sensor fusion approaches that explicitly handle failure modes [20]. Planning and control are complicated by occlusion and deformability: Robots act under incomplete state information and plant motion induced by wind or by tool interaction. Safety and human–robot interaction impose additional constraints, and safety standards for autonomous agricultural machinery emphasise formal risk assessment and verification/validation beyond task success alone [21,22].

Prior surveys provide valuable catalogues of systems and enabling methods, but they often remain organised by a single dominant lens, which can limit interpretability across platforms and deployment settings. Reviews structured primarily by sensing modality or by application lists may understate how platform constraints shape feasible autonomy and how evaluation setting conditions generalisability, while subdomain-focused reviews, such as those centred on UAV applications or harvesting perception, provide limited cross-task comparison of design tradeoffs and evaluation practices [9,15,17,23]. Building on this literature, the present review makes four specific contributions. First, it organises the field jointly through task, platform, autonomy-stack, and evaluation-setting lenses. Second, it maps each task onto a sense–decide–act pipeline to expose bottlenecks and failure modes across autonomy modules. Third, it synthesises evaluation metrics, protocol design, benchmark progression, and reporting guidance to improve comparability and reproducibility. Fourth, it strengthens the engineering perspective of this review by explicitly incorporating implement/tool specification and mechanical reliability into the interpretation of actuation outcomes in weeding, spraying, and harvesting. It is worth noting that, unless otherwise stated, the conceptual figures and synthesis tables in this review are author-generated summaries based on the reviewed literature, with direct source citations provided in the main text and literature-summary tables where appropriate.

This review is closely aligned with the United Nations Sustainable Development Goals (SDGs), particularly SDG 2 (Zero Hunger), which calls for improved agricultural productivity and sustainable food production systems, and SDG 12 (Responsible Consumption and Production), which emphasises efficient resource use and the reduction in food losses and chemical waste [24]. By examining robotics in precision agriculture through both task-based and platform-based perspectives, this paper highlights how monitoring, weeding, spraying, and harvesting robots can contribute to these goals through more targeted interventions, reduced input overuse, and improved operational efficiency [25]. In addition, this review is relevant to SDG 13 (Climate Action), because agricultural robotics has been identified as a pathway toward net-zero agriculture through improved nitrogen-use efficiency, reduced waste, electrified robotic vehicles, and lower environmental impact per unit of production [26].

This paper first defines the review methodology, taxonomy, and coding scheme. It then synthesises platform archetypes and task-centric autonomy pipelines, followed by cross-cutting enabling technologies. A dedicated section analyses evaluation and benchmarking practices and proposes a reporting checklist. This paper concludes with a synthesis of design tradeoffs, a decision framework for selecting platform/toolchains, and a staged research agenda.

2. Review Methodology and Taxonomy

2.1. Search Strategy

The literature search was designed to capture the interdisciplinary nature of agricultural robotics, where relevant work is disseminated across robotics, agricultural engineering, sensing, and remote-sensing venues. To balance coverage and precision, we queried multidisciplinary indexing services (e.g., Web of Science and Scopus) alongside robotics-focused digital libraries (e.g., IEEE Xplore, MDPI, and the ACM Digital Library). In addition, scholar-indexed search engines were used as a complementary step to improve recall in cases where terminology varies across communities or where relevant studies are inconsistently indexed. The search covered the literature indexed up to 5 March 2026, and no explicit lower bound on publication year was imposed.

Search queries were formulated by combining three term groups. The first group targeted agricultural tasks: monitoring (including scouting and phenotyping), weeding (including weed control), spraying (including plant protection and variable-rate application), and harvesting (including picking). The second group targeted robotic platforms: unmanned ground vehicles (UGVs) and field robots, unmanned aerial vehicles (UAVs/drone), gantry/rail/cable-driven systems, and greenhouse robots. The third group targeted autonomy and enabling technologies: simultaneous localisation and mapping (SLAM), perception for detection/segmentation, motion planning, and compliant control/manipulation. Boolean operators and field restrictions (title/abstract/keywords) were applied where supported, and query variants were used to account for common synonyms and spelling differences. This review was scoped to crop-production precision-agriculture robotics in arable and horticultural systems; livestock robotics (e.g., milking, feeding, barn inspection, and animal monitoring) was excluded because its task structure, sensing environment, welfare and biosecurity requirements, and evaluation protocols differ substantially from those of crop-field robotics and therefore warrant a separate dedicated synthesis.

Study selection followed a transparent, staged screening process. After consolidating retrieved records and removing duplicates, titles and abstracts were screened for relevance to agricultural robotic systems, addressing at least one target task and including substantive technical content (e.g., platform description, sensing/perception, navigation, planning/control, actuation/tooling, or field evaluation). Full-text screening was then conducted to confirm eligibility and extract data for synthesis. Figure 1 summarises the literature selection workflow using a PRISMA-style flow diagram reporting records identified, duplicates removed, records screened, full texts assessed, and exclusions with reasons, consistent with PRISMA 2020 guidance [27].

2.2. Taxonomy Definition

To ensure consistent categorization across heterogeneous studies, this review adopts an explicit taxonomy along four axes: task, platform, autonomy stack, and evaluation setting. The task categories are defined as follows. Monitoring covers robotic sensing, mapping, and estimation of crop and field state to support agronomic decision-making. Weeding encompasses selective weed suppression or removal using mechanical, thermal, or chemical micro-dosing interventions. Spraying includes controlled deposition of plant protection products or other agrochemicals, with emphasis on variable-rate application, targeting accuracy, and drift mitigation. Harvesting refers to robotic picking and handling operations that require integrated perception-to-grasp pipelines and strategies to preserve produce quality.

Robotic platforms are categorized as unmanned ground vehicles (UGVs), unmanned aerial vehicles (UAVs), gantry/fixed-structure systems, greenhouse robots, and hybrid or multi-robot systems combining two or more platform types and/or coordinated teams. The autonomy stack is decomposed into the functional components most frequently reported in the literature: perception, localisation, planning, control, actuation/tooling, safety, and human–robot interaction (HRI). Evaluation context is coded independently of platform and task to avoid conflating capability with the test environment; settings are classified as lab/bench, greenhouse, open-field (single season), and open-field (multi-season or multi-site).

Table 1. Task, platform, autonomy-stack, and evaluation-setting taxonomy used in this review, together with the corresponding definitions and typical sensors/implements. The taxonomy structure is aligned with prior agricultural robotics and sensing-in-agriculture reviews like [15,20].

Category	Definition	Typical Sensors/Implements
Tasks
Monitoring	Robotic data acquisition and interpretation for crop/soil status (e.g., phenotyping, stress/disease scouting, mapping), with outputs such as maps, indices, counts, or alerts.	Sensor suite (RGB/multispectral/hyperspectral/thermal/ LiDAR, GNSS/IMU, soil probes), mounting (mast/boom/gimbal), and any sampling/end-effector used for in situ measurements.
Weeding	Autonomous or semi-autonomous weed detection and removal, including intra-row/inter-row operations with mechanical, thermal, electrical, or chemical means.	Crop/weed sensing (RGB, multispectral, depth/LiDAR), implements (mechanical hoe/cultivator, rotary tools, laser/thermal, micro-sprayer/nozzle for spot herbicide), and actuation modality.
Spraying	Targeted or blanket application of agrochemicals (pesticide, herbicide, fertilizer) via robotic platforms, emphasizing dosage control, drift mitigation, and coverage uniformity.	Nozzle type, pump/valve control, droplet/flow sensors if any, perception for target selection, boom configuration, UAV payload specs (if applicable).
Harvesting	Autonomous identification, approach, detachment, and handling of produce (or selective picking), including manipulation in cluttered canopies and damage-avoidance constraints.	Fruit/produce perception (RGB/depth/LiDAR), gripper/end-effector type (soft/compliant, suction, cutting tool), force/torque sensing, and any conveyors/bins.
Platforms
UGV	Ground-based robot operating in rows/alleys, typically constrained by terrain, canopy clearance, and turning radius; may carry implements or manipulators.	Wheel/tracked locomotion, steering type, onboard sensors, implement mounting (front/rear/underbody), power system.
UAV	Aerial robot used for sensing and/or spraying, characterized by limited payload/flight time and strong dependence on environmental conditions.	Camera payload (RGB/multispectral/thermal), RTK GNSS, gimbal, spray kit (tank/nozzles), and coverage planning tools.
Gantry/fixed-structure	Robots constrained to rails, frames, or fixed infrastructure (field gantries, overhead rails), offering repeatability and stable sensing at the cost of limited workspace.	Rail/actuator type, sensing modality, repeatability specs, and end-effector integration.
Greenhouse robot	Robots designed for GNSS-denied, structured indoor cultivation with narrow aisles and high worker proximity; often integrated with greenhouse infrastructure.	Localisation aids (fiducials/UWB/SLAM), perception sensors, and task implements; safety sensors (LiDAR, bumpers).
Hybrid/multi-robot	Systems combining multiple platforms (e.g., UAV + UGV) or coordinated robot teams; emphasis on task allocation, data fusion, and fleet management.	Coordination method, comms stack, shared maps, heterogeneous sensing, and any docking/charging logistics.
Autonomy stack modules
Perception	Sensing and interpretation (detection/segmentation/tracking, state estimation from raw sensors) for crops, weeds, obstacles, and targets.	Sensor modalities, model class (classical vs. DL), domain adaptation strategy, and runtime constraints.
Localisation	Pose estimation in the farm environment using GNSS/RTK, SLAM, visual odometry, UWB, fiducials, or map-based methods.	Sensors used (GNSS/IMU/LiDAR/camera/UWB), map representation, and failure cases (occlusion/GNSS outage).
Planning	Decision-making for coverage, motion, manipulation sequencing, and task allocation under constraints (time, energy, safety, crop damage).	Planner type (coverage/path/task), constraints modeled, re-planning triggers, and multi-robot coordination.
Control	Low-level control for vehicle dynamics and manipulator control (tracking, force/impedance), including disturbance rejection under terrain/canopy interactions.	Control architecture, feedback signals, control rates, and stability/robustness evidence.
Actuation	Physical interaction with crop/weed/chemical delivery: end-effectors, sprayers, cutters, grippers, and mechanisms for handling.	Actuator type, end-effector design, compliance, force limits, and maintenance/cleaning considerations.
Safety	Measures ensuring safe operation around workers and crops (collision avoidance, safe stop, redundancy, functional safety concepts).	Safety sensors, protective zones, stop logic, standards referenced (if any), and incident reporting.
HRI	Human–robot interaction for supervision, teleoperation, shared autonomy, interfaces, alerts, and workflow integration.	Interface type (tablet/voice/teleop), autonomy level, operator workload measures, and training requirements.
Evaluation settings
Lab/bench	Controlled tests of sensing, control, actuation, or spray/coverage rigs with repeatable conditions and limited environmental variability.	Test rigs, calibration targets, artificial plants/rows, and sensor calibration procedure.
Greenhouse	Semi-structured cultivation with aisle constraints, GNSS denial, controlled but variable lighting/humidity, and high worker proximity.	Localisation aids, lighting control, and safety measures used.
Open-field single season	Outdoor evaluation within one growing season and one site, with limited capture of seasonal/domain variability.	Crop type, row geometry, soil/terrain, and weather reporting.
Open-field multi-season/multi-site	Outdoor evaluation spanning multiple seasons and/or sites to capture domain shift (cultivar, weather, management practice, sensor ageing).	Site/season count, domain shift factors, and generalisation protocol.

summarises the scope and taxonomy of this review and defines the categories used to code the selected studies, including task type, platform class, autonomy-stack component, and evaluation setting. This taxonomy is aligned with conventions used in prior agricultural robotics and sensing-in-agriculture reviews that separate application classes, mobility/platform types, and sensing modalities while additionally enabling explicit mapping of each study to autonomy components for cross-task synthesis [15,20].

A structured coding scheme was applied to support consistent comparison across studies, given that claims of “autonomy” vary widely in operator involvement, sensing redundancy, and safety constraints. For each included paper, we extracted and coded the crop and operating environment (crop type; greenhouse versus open-field; terrain and canopy characteristics where reported), the sensor suite (e.g., RGB/multispectral/thermal imaging, LiDAR, GNSS/RTK, IMU, force/torque), the localisation method (e.g., GNSS/RTK, LiDAR- or vision-based SLAM, fiducial-based, map-assisted), the perception approach (e.g., classical vision, CNN-based detection/segmentation, transformer-based models, multimodal fusion), the implemented tool or end-effector (e.g., mechanical/thermal/chemical weeding implement, sprayer/nozzle configuration, gripper/cutter type and compliance strategy), the evidenced autonomy level (teleoperation, supervised autonomy, mission-level autonomy, or fully autonomous as supported by the described protocol), and the evaluation protocol and metrics (including dataset split design where applicable, field-trial procedures, success criteria, throughput, accuracy, crop damage, chemical reduction, energy use, and safety-related outcomes). Openness was coded separately to record the availability of data, code, models, and implementation details sufficient for replication.

To avoid success-only summarisation, each record included a dedicated limitations field capturing explicitly stated constraints and failure modes (e.g., illumination sensitivity, occlusion, seasonal domain shift, localisation loss, actuator clogging, bruising, safety stop triggers). This enables mechanistic synthesis by linking observed performance to operating conditions and system design choices. In line with robustness-focused crop imaging reviews that highlight the influence of acquisition protocol, metadata, and evaluation split design on reported performance, these elements were explicitly captured to support later assessment of generalisability [19].

2.3. Autonomy Levels

In this review, autonomy is defined operationally in terms of operator involvement across the perception–planning–safety loop rather than by informal labels (e.g., fully autonomous). Specifically, autonomy level is characterised by which functions require human initiation, supervision, or intervention and under what conditions responsibility is transferred back to the operator (e.g., perception failure, localisation degradation, safety triggers). This framing is important because safety guidance for automated agricultural machinery typically emphasises not only design principles but also verification and validation activities spanning partially automated through autonomous operation; therefore, any autonomy claim should be coupled to the presence and testing of safety functions (e.g., hazard detection, safe-stop behaviour, fault handling) and to evidence demonstrating their performance in representative conditions. In parallel, evaluation rules used in field-robot competitions provide a useful operational lens on autonomy by enforcing constraints that mirror deployment realities, such as limits on external localisation aids, mission time bounds, and penalties for crop damage, thereby clarifying what “autonomous” entails in practice when navigation, task execution, and safety must be achieved under explicit rules and measurable outcomes [9,22]. The autonomy levels used in this review are summarised in

Table 2. Operational autonomy levels for agricultural robots.

Autonomy Level	Operator Involvement	Perception	Planning	Safety
Teleoperated	Continuous human control of motion and action	Human interprets scenes; robot provides sensor feed	Human plans trajectories/actions	Manual E-stop; human ensures separation
Supervised autonomy (task-level)	Human initiates tasks; intervenes on failure	Robot detects relevant objects in nominal cases	Robot executes nominal plans; operator recovers exceptions	Basic E-stop and stop-on-command
Conditional autonomy (exceptions handled by operator)	Robot runs nominally; operator resolves edge cases	Robust in known conditions; uncertainty flagged	Nominal planning with limited replanning	Safety logic present but limited validation
High autonomy (common exceptions handled)	Operator monitors; rare intervention	Robust to moderate shift; redundancy/fusion used	Online replanning and recovery behaviours	Explicit risk models; validated stop behaviours
Field autonomy with human presence	Operates near workers under protocols	Perception includes humans/obstacles	Plans include safety buffers and emergency behaviours	Safety mechanisms tested under protocols
Unattended autonomy (restricted domain)	Remote monitoring only; no on-site operator	High robustness in restricted environments	Full mission autonomy in defined domain	Formal verification/validation evidence

, ranging from fully teleoperated systems to unattended autonomy in restricted domains supported by verification and validation. Intermediate levels capture increasing perception robustness, planning adaptability, exception handling, and safety assurance, enabling consistent comparison across systems and deployment contexts.

3. Platforms in Agricultural Robotics

Platform choice shapes the entire autonomy stack because it constrains sensing geometry, computational and energy budgets, tool integration, and safety envelopes. Surveys of field robots commonly distinguish aerial platforms for broad monitoring from ground platforms for close-range sensing and physical intervention, while phenotyping-system reviews further highlight the prevalence of rail, gantry, and other fixed-structure alternatives where repeatability is prioritised [8,10].

These platform archetypes, summarised in Figure 2, are defined in this review by mobility constraints, sensing/actuation viewpoint, and infrastructure dependence. In this sense, aerial platforms support broad overhead sensing, ground platforms support close-range sensing and physical intervention, and fixed-structure systems prioritise repeatable sensor pose, consistent with distinctions commonly used in prior agricultural robotics and phenotyping-system reviews [15,20]. Greenhouse and hybrid systems are included as distinct classes because they introduce GNSS-denied constrained operation and coordinated aerial–ground workflows, respectively, which carry different autonomy and evaluation implications [28].

Accordingly, this section examines unmanned ground vehicles (UGVs), unmanned aerial vehicles (UAVs), gantry and fixed-structure systems, greenhouse robots, and hybrid configurations, with emphasis on how each platform class shapes sensing and actuation capabilities, localisation and planning requirements, and safety constraints. This section concludes by discussing hybrid systems and multi-robot coordination, where complementary aerial–ground workflows can improve coverage and operational efficiency but introduce additional challenges in task allocation, communication, and system-level reliability. In order to standardize comparisons across studies, the evaluation metrics and reporting items are summarized in

Table 3. Comparison of platform archetypes in agricultural robotics.

Platform	UGV	UAV	Gantry/Fixed-Structure	Greenhouse Robot	Hybrid
Mobility	High within farm terrain; constrained by row spacing/soil	Very high spatial coverage	Low (fixed infrastructure)	Moderate; structured aisles	High (combined mobility modes)
Payload	High; can carry tools and supplies	Low–moderate; sensor-limited	Moderate–high (mounted on structure)	Moderate	Distributed across platforms
Energy endurance	Medium to high	Low, 10–90 min	Very high	High to very high	Distributed
Localisation reliability	Depends on GNSS/RTK and/or SLAM; canopy and soil effects	Often GNSS-based; sensitive to wind and RTK availability	High repeatability (fixed reference frame)	SLAM/fiducials; GNSS-denied	Cross-platform frame alignment required
Operational cost	Medium (51% lower operational cost per hour than the tractor [29])	Medium to high (e.g., monitoring costs USD 45–60 per hectare per year [30])	High infrastructure cost, low-to-medium operational cost	Medium operational cost.	Higher system-integration/ operational cost
Scalability	Moderate; redeployable	High for monitoring deployments	Low–moderate	Moderate in similar facilities	Potentially high
Safety	Higher near humans and tools	Moderate; airspace constraints	Lower in restricted operational design domain	High in confined spaces	Coordination risk
Typical tasks	Weeding, spraying, under-canopy monitoring, some harvesting	Monitoring, mapping, scouting	Monitoring, controlled interventions	Monitoring, harvesting, targeted treatments	Monitoring + intervention pipelines
Refs.	[9,29,31,32,33]	[17,30,34,35,36,37]	[16,38,39]	[39,40,41,42,43,44]	[30,39,45,46]

Note: Quantitative entries are literature-reported representative values. Unless reported in normalized units, they should be interpreted with reference to the original benchmark, payload, working width, and operating conditions.

.

Table 3 serves as a reference for the quantitative and qualitative criteria used throughout this review. The presented quantitative values are literature-reported representative figures, not results from a single standardised benchmark. Therefore, cost and endurance metrics should be interpreted in the context of the original study conditions and assumptions. These results are supported by both synthesis reviews and representative experimental or field-evaluation studies for the major platform classes. In addition to synthesis reviews, Table 3 is explicitly anchored in representative primary platform studies, including UGV mapping and phenotyping systems [32,33], UAV field surveying and variable-rate spraying systems [34,35], and a hybrid UAV–UGV coordination workflow validated in orchard conditions [45]. For the fixed-structure and greenhouse classes, the table is also consistent with representative primary systems already discussed in this section, including automated field phenotyping and greenhouse navigation platforms.

To improve cross-study comparability, quantitative platform indicators should be expressed, where possible, using both absolute and normalized units. Operational cost should be reported as currency per hour and, for area-based operations, additionally as currency per hectare together with effective field capacity (ha h⁻¹) and the included cost components. Energy endurance should be reported as operating time per charge or fuel tank (min or h) under stated payloads and duty cycles and complemented by an area- or task-normalized intensity metric, such as kWh ha⁻¹ for electric platforms or L ha⁻¹ for fuel-based systems. Relative claims (e.g., lower cost or longer endurance) should also identify the comparator, task, and operating conditions to avoid overinterpretation of the literature values reported under different assumptions [9,22,29].

Environmental conditions strongly shape feasible platform designs and limit the direct comparability of platform-level performance across studies. Open-field arable systems typically favour larger platforms with greater endurance and working width to cover large areas efficiently, although these features may increase soil compaction and reduce manoeuvrability. In contrast, horticultural and orchard environments impose tighter geometric constraints due to row spacing, canopy structure, and crop proximity, thereby increasing the importance of compact designs, suitable sensor viewpoints, and safer, more compliant tools. Greenhouse systems add further constraints, including GNSS-denied operation, narrow aisles, worker proximity, and exposure to humidity and condensation, which often favour small platforms and infrastructure-assisted navigation. These environment-dependent constraints should therefore be considered part of the operational design domain when comparing platform classes across studies.

3.1. Unmanned Ground Vehicles

UGVs commonly use wheeled or tracked locomotion to trade off stability on uneven soil against manoeuvrability within narrow row corridors. Their operation is strongly constrained by crop geometry, as row spacing and canopy closure evolve across growth stages and thereby limit the admissible vehicle envelope while shaping sensor viewpoint and coverage. Performance is further influenced by terrain–soil interaction, where wheel slip and sinkage degrade localisation accuracy and increase energy demand, and platform vibration reduces perception stability and calibration fidelity. Consistent with these constraints, navigation-oriented evaluation settings explicitly penalise crop damage and emphasise robust row-following and recovery behaviours under conditions such as missing plants and irregular row structure [9].

3.2. Unmanned Aerial Vehicles

UAVs provide rapid field coverage and a top–down sensing geometry that is well suited for monitoring and mapping, most commonly using red-green-blue (RGB) and multispectral imaging. UAV-focused reviews report that practical deployments are often limited by flight endurance, payload capacity, and calibration procedures required to maintain consistent radiometric and geometric outputs across repeated missions [17].

Operational feasibility is additionally shaped by aviation regulations. In the European Union (EU) open category, the European Union Aviation Safety Agency (EASA) specifies constraints such as operation within visual line of sight (VLOS) and a nominal 120 m height limit, while in the United States, the Federal Aviation Administration (FAA) outlines Part 107 requirements, including keeping the UAV within sight and other operational limitations [36,37]. Collectively, these constraints restrict routine beyond visual line of sight (BVLOS) operation and high altitude mapping, and they therefore influence whether UAV outputs are used for near real-time decisions or as offline decision support products.

3.3. Gantry and Fixed-Structure Systems

Gantry and rail systems typically sacrifice mobility to achieve high positional accuracy and measurement repeatability. The Field Scanalyzer rail platform is a representative example, as it uses fixed infrastructure to translate a multisensor payload over experimental plots with controlled pose, enabling high-throughput repeated measurements and dense temporal sampling across the crop lifecycle [16]. Related fixed-structure and cable-driven configurations are also reported in greenhouse and urban farming contexts, where suspended architectures increase workspace coverage while maintaining end effector pose control and reducing interaction with the ground [38]. The main limitation is scalability and cost, since the required infrastructure restricts deployment to instrumented plots or dedicated facilities, and outcomes may not readily generalise to broader farm environments without comparable structure [16].

3.4. Greenhouse Robots

Greenhouse robots operate in GNSS-denied, spatially constrained, worker-proximate environments with partial structural regularities such as aisles, benches, and, in some cases, rails. This distinguishes them from open-field and orchard robots, which typically operate over larger and less confined spaces and more often rely on GNSS-supported navigation. Consequently, greenhouse navigation studies commonly adopt light detection and ranging (LiDAR)- and vision-based mapping and localisation pipelines while placing greater emphasis on confined-space manoeuvring, obstacle avoidance, and safe operation in close proximity to workers, sometimes exploiting environmental features such as rails for guidance [42,43]. Greenhouse deployment also introduces sensing artefacts, including condensation and specular reflections, and requires hardware resilience to elevated humidity and potential chemical exposure [42].

3.5. Hybrid Systems and Multi-Robot Coordination

Hybrid agricultural systems frequently pair unmanned aerial vehicles (UAVs) for field-scale mapping with unmanned ground vehicles (UGVs) for localised intervention, enabling a monitor-then-act workflow in which aerial products guide targeted ground operations. Reviews of integrated workflows identify calibration and cross-sensor harmonisation as persistent bottlenecks and note that the perception of execution pipelines often relies on interoperability with farm information systems to translate maps into actionable prescriptions [46]. This coordination layer deserves more explicit treatment, because agricultural UAV–UGV systems differ not only in sensing and actuation roles but also in how aerial and ground decisions are coupled.

Multi-robot coordination further imposes requirements on communication reliability and consistent shared localisation frames; in practice, agricultural deployments more often adopt division of labour and loosely coupled task allocation rather than tightly coupled formation control, reflecting limitations in connectivity, as well as the need for conservative safety assurance in shared workspaces [46]. UAV–UGV coordination architectures can be organised as centralised, decentralised, or hybrid/hierarchical. In centralised coordination, a supervisory layer such as a farm management information system aggregates aerial and ground observations to compute shared task allocations or routes, improving global consistency at the cost of infrastructure dependence and possible communication latency [47,48]. In decentralised coordination, each platform retains greater local decision authority and exchanges only mission-critical state information, which can improve responsiveness and scalability but may reduce global optimality and shared situational awareness [47,48]. Hybrid schemes combine these principles using shared map-level or mission-level planning together with onboard local autonomy, and they are therefore often the most realistic pattern for agricultural deployments [28,49].

Agriculture-specific studies are also moving beyond simple map handoff toward proactive coordination, in which UAV observations are converted into georeferenced exclusion zones or route constraints before UGV execution; recent orchard experiments show that such centralised UAV–UGV planning can reduce detours and support safer navigation in dynamic environments [50]. The same monitor–plan–act logic is also relevant to targeted harvesting and other selective interventions, where aerial scouting can prioritise blocks, rows, or candidate targets for subsequent ground manipulation, although fully integrated UAV–UGV harvesting workflows remain an emerging direction rather than a mature benchmarked application [28].

4. Task-Centric Review

In this task-centric review, the literature is organised around four core tasks in agricultural robotics: monitoring and phenotyping, weeding, precision spraying, and harvesting. These task categories follow the taxonomy defined in Section 2.2 and summarised in Table 1, which is used consistently throughout the remainder of this manuscript. For each task, an end-to-end autonomy pipeline is considered, from sensing and calibration to perception, localisation, safety-constrained planning, actuation, and outcome logging. The four task pipelines are summarised in Figure 3 to clarify recurring bottlenecks and to support consistent interpretation of reported evaluation evidence. Figure 3 is a conceptual synthesis by the authors, derived from representative task-pipeline patterns reported in prior studies on robotic monitoring and phenotyping, robotic weeding, precision spraying, and robotic harvesting [32,51,52,53]. Throughout Section 4, the discussion follows the four-task taxonomy adopted in this review, namely monitoring, weeding, spraying, and harvesting, as defined in Table 1. For clarity, intervention tasks are additionally classified by actuation modality and targeting granularity, as summarized in

Table 4. Representative weeding approach categories and reported evaluation characteristics.

Attribute	Inter-Row Mechanical Cultivation	Intra-Row Mechanical Micro-Cultivation	Laser/Thermal Spot Weeding	Chemical Micro-Dosing
Typical sensors	RGB or LiDAR for row detection	RGB/RGB-D	RGB/RGB-D	RGB
Typical model/control	Row following + tool control	Crop–weed segmentation + precise actuation	Detection + pointing control	Detection + nozzle timing
Target characterisation	Row centreline/inter-row weed band; weed pressure or density along the row	Crop/weed position within the row; intra-row spacing; protected vs. treatment zones	Weed stem/apical target location; weed species/growth stage/size; safe treatment zone	Weed class or patch extent; target footprint; weed density/coverage; crop spacing and off-target exclusion zone
Implement/tool	Hoe/tines	In-row devices (e.g., finger weeder, torsion weeder, in-row knife/blade, rotating star/disc/brush tools; actuator timing and crop-protection zone)	Laser module (report wavelength, power, spot diameter, standoff distance, exposure time/dose)	Fast nozzles/DoD
Reported strengths	Robust without plant-level ID	Chemical-free near crops	Soil disturbance reduced	Reduced chemical use
Reported weaknesses	Limited intra-row control	High precision required	Safety constraints; energy/throughput limits	Synchronisation errors; drift
Evaluation setting	Open field	Open field	Open field	Open field
Metrics used	Weed reduction; crop damage	Plant-level success; crop injury	Weed suppression; safety reporting	Hit rate; off-target risk
Refs.	[54,55,56]	[57,58,59]	[60,61,62,63]	[64,65,66]

and

Table 5. Representative spraying approach categories and reported evaluation characteristics.

Attribute	Map-Based Variable-Rate	Real-Time Canopy-Sensed Variable-Rate	Plant-Level Micro-Spraying
Typical sensors	UAV RGB/multispectral maps + GNSS/RTK	LiDAR/ultrasound + GNSS/IMU	RGB/RGB-D vision + encoder/IMU
Typical model/control	Prescription map + section/nozzle control	Canopy sensing + flow/PWM control	Detection/segmentation + nozzle gating
Implement/tool	Boom/orchard sprayer (nozzle set, pressure, flow rate)	Variable-rate nozzles/air-assist (pressure, flow, sensor–nozzle offset)	Micro-nozzles/micro-jets (spot size, pressure, response time)
Reported strengths	Integrates with farm planning	Responsive dosing	High selectivity
Reported weaknesses	Map errors propagate	Turbulence and wind coupling	Latency and occlusion failures
Evaluation setting	Open field/orchard	Orchard	Open field
Metrics used	Input reduction (%); deposition/coverage; field capacity (ha h−1)	Deposition uniformity; drift proxy/loss; chemical saving (%)	Target hit (%); off-target deposition (%); latency (ms)
Refs.	[35,67,68,69]	[53,70,71,72]	[66,73,74,75]

. Specifically, Table 4 groups weeding systems into inter-row and in-row mechanical cultivation, thermal or laser spot treatment, and chemical micro-dosing, while Table 5 groups spraying systems into map-based variable-rate application, real-time canopy-sensed variable-rate application, and plant-level micro-spraying.

4.1. Monitoring and Phenotyping

In this review, monitoring refers to robotic sensing, mapping, and estimation of crop or field state to support agronomic decision-making, including scouting and phenotyping tasks as listed in Table 1. Monitoring and phenotyping robots aim to estimate crop and field state with sufficient spatial registration to inform management decisions, with deployment contexts ranging from breeding trials that prioritize repeatability and dense temporal sampling to production fields where coverage and timeliness dominate. Unmanned aerial vehicles (UAVs) are widely used for monitoring through red-green-blue (RGB) and multispectral imagery, producing trait and anomaly maps that are often evaluated as spatial segmentation problems rather than through downstream actuation outcomes. Large annotated datasets such as Agriculture-Vision exemplify this paradigm, enabling assessment using overlap-based metrics that quantify the spatial localisation of field anomalies [76].

In contrast, unmanned ground vehicles (UGVs) extend monitoring under canopy and support additional sensing modalities, including RGB-depth (RGB-D) cameras and light detection and ranging (LiDAR), to capture plant structure traits at close range. Field phenotyping reviews further indicate that UGV operation is strongly constrained by crop geometry and growth stage, which jointly affect traversability and sensing viewpoint [10]. Multi-site deployments of small under-canopy phenotyping robots have also been reported in breeding settings, highlighting the potential of repeated measurements while exposing practical challenges in reliability and user workflow during sustained field operation [76].

Sensors for monitoring are typically selected to maximise information content while meeting field constraints such as cost, power, and ease of deployment. Multispectral cameras support vegetation indices and stress indicators, light detection and ranging (LiDAR) supports three-dimensional reconstruction and canopy structure measurement, and red-green-blue depth (RGB-D) sensing can provide local three dimensional perception under canopy. In plant phenotyping, multispectral imagery is commonly reduced to vegetation indices that represent greenness, chlorophyll content, canopy structure, or stress. Widely used examples include NDVI for greenness and biomass, NDRE and GNDVI for chlorophyll-related traits, EVI for dense canopies, SAVI and MSAVI when the soil background is visible, and PRI for physiological stress assessment. Because these indices remain sensitive to illumination conditions and sensor calibration, reliable phenotyping depends on radiometric correction and consistent image acquisition settings.

For georeferencing, monitoring platforms often rely on the global navigation satellite system (GNSS) with real-time kinematic (RTK) corrections; however, in under-canopy or protected environments, navigation commonly shifts toward simultaneous localisation and mapping (SLAM) combined with sensor fusion. A key challenge is the interaction between domain shift and calibration: Changes in illumination, season, and acquisition settings alter radiometric and geometric properties, which can cause learned trait estimators to drift without clear failure signs. Robustness-focused reviews therefore recommend evaluation protocols that explicitly test spatial and temporal generalisation—for example, by holding out entire farms or sites and seasons—to support credible performance claims [18].

Evaluation metrics commonly include trait estimation errors relative to manual ground truth, consistency across runs, and coverage/time proxies, but reported protocols vary in how they describe acquisition timing, environmental conditions, and calibration. A recurring omission is validation of decision impact: monitoring studies often demonstrate sensing accuracy but do not test whether the produced maps or trait estimates lead to improved intervention planning under farm constraints [16].

Table 6. Representative approaches in agricultural robotic sensing and phenotyping.

Attribute	UAV Multispectral Anomaly Mapping	UGV Under-Canopy Phenotyping	Fixed-Structure Gantry Phenotyping
Typical sensors	RGB & NIR multispectral	RGB-D; LiDAR; RGB	Hyperspectral; thermal; LiDAR; RGB
Typical model/control	Segmentation/classification & georeferencing	Trait extraction; 3D reconstruction	Repeatable time-series acquisition
Typical outputs/traits	NDVI/NDRE and related vegetation indices; stress/anomaly maps; LAI/SPAD chlorophyll and yield proxies where validated	Plant counts/plant centres; canopy height; crown width/volume; local 3D structural traits	Repeatable time-series of canopy height, temperature, spectral traits, growth trajectories, and stress progression
Implement/tool	None	None or light probes	Rail/gantry
Reported strengths	High coverage; consistent survey geometry	Close-range structural sensing	High repeatability
Reported weaknesses	Radiometric calibration; domain shift sensitivity	Traversal limited by canopy geometry	Infrastructure cost; limited scalability
Evaluation setting	Open field	Open field	Field facility
Metrics used	mIoU; precision–recall; map consistency	Trait error vs. ground truth	Repeatability metrics; trait tracking stability
Refs.	[76,79,80,81]	[10,32,33,77]	[16,82,83]

summarises representative agricultural robotic sensing and phenotyping approaches, together with their sensing modalities, technical characteristics, and evaluation contexts. To make the phenotyping and modelling content more explicit, Table 6 also identifies representative outputs and traits reported by these systems, including vegetation indices such as NDVI and NDRE, physiological proxies such as leaf area index (LAI) and chlorophyll-related measures, and structural traits such as plant counts, canopy height, and canopy width/volume [32,77,78].

4.2. Weeding

Weeding robots aim to suppress weeds while preventing crop injury, particularly in row crops and high-value systems where labour shortages and reduced herbicide use are important drivers. The literature commonly distinguishes three main approaches: mechanical weeding using intra-row or inter-row cultivation tools; thermal weeding with an emphasis on laser-based targeting; and chemical micro-dosing or spot spraying. These modalities impose different autonomy requirements. Mechanical tools operating close to crops require high-precision localisation and robust crop weed discrimination, whereas inter-row tools can often rely more on row detection. Laser systems demand accurate pointing together with stringent safety mechanisms, while micro-dosing requires tightly synchronised perception to nozzle timing [84].

Sensors for weeding are frequently RGB cameras due to plant-scale resolution needs; RGB-D and LiDAR are added when depth or structure improves separation. Public datasets recorded from field robots over growth stages, including sugar beet datasets spanning plant emergence through later stages, provide benchmarks for crop/weed segmentation and mapping and illustrate how occlusion and growth stage complicate discrimination [85]. Laser-weeding surveys discuss the dependence of effectiveness on crop spacing, weed density, and safety design, indicating that performance claims must be contextualised by operational conditions and actuation constraints [85].

From a mechanical-design perspective, the engineering relevance of laser-weeding modules depends on reporting actuator-level parameters rather than only naming the modality. In particular, laser type and wavelength, output power, spot diameter at the target plane, standoff distance, exposure time, and delivered energy or energy density should be specified, since these quantities govern treatment selectivity and weed-destruction effectiveness. Representative recent systems range from a 5 W, 450 nm diode laser with an approximately 4 mm beam width at the stem and 2 s exposure to a 50 W [51], 1940 nm thulium fiber laser with an approximately 1.5 mm beam diameter [60], while review synthesis indicates spot diameters of roughly 0.6–5 mm and power levels of 4–50 W across different laser types [51,60,86]. Reporting these parameters would improve the engineering interpretability and reproducibility of laser-weeding studies.

Localisation and navigation in weeding robots typically exploit row structure, but performance can degrade with missing plants, non-parallel rows, and shadow-induced perception loss. Reported evaluation metrics therefore vary widely, including weed removal or kill rate, proxies for crop damage, area-level productivity measures, and reductions in chemical or other inputs. A key system bottleneck is the tight coupling between perception uncertainty and irreversible actuation: Because weeding tools can injure crops, uncertainty often forces conservative decision thresholds, which reduces weed removal efficacy or operating speed. This creates a precision–throughput–crop-safety tradeoff that must be addressed at the system level rather than by perception accuracy alone [84]. Table 4 summarises representative weeding approach categories and their reported evaluation characteristics, including typical sensing modalities, control strategies, implements, operational strengths and weaknesses, and commonly used performance metrics. To make intervention-relevant state variables more explicit, Table 4 also identifies representative target-characterisation variables that condition robotic weeding performance, including weed species or functional group, growth stage, crop/weed position, protected zones, and density- or coverage-based quantities used to define treatment opportunity [51,73,87,88].

4.3. Spraying

Precision spraying aims to achieve effective deposition on target plant structures while reducing off-target drift and over-application, with important differences between orchard or vine systems and arable crops. In this review, spraying robots are classified into three categories, as summarized in Table 5: (i) map-based variable-rate application, typically based on prescription maps derived from UAV or remote sensing data; (ii) real-time canopy-sensed variable-rate spraying, which relies on canopy measurements, such as LiDAR or ultrasonic sensing, coupled with PWM or flow control; and (iii) plant-level micro-spraying or spot spraying, which is typically enabled by vision-based nozzle control. Relevant standards define field measurement procedures for spray distribution in tree and bush crops and for drift assessment, emphasizing that deposition quality is inherently distributional and that meteorological conditions and measurement distances should be recorded to enable meaningful comparison across trials [89,90].

Variable-rate spraying systems commonly combine canopy sensing—for example, using light detection and ranging (LiDAR) or ultrasonic ranging—with flow and nozzle control. This couples perception of canopy presence or volume with localisation along rows, planning of dosage schedules that account for actuation delays, and execution through nozzle gating, droplet control, and, in some cases, shielding. Recent orchard studies report LiDAR-based variable-rate sprayers and corresponding evaluation approaches, typically emphasizing deposition distributions and drift-related proxies, although the completeness of reported protocols varies across the literature [53,71].

Sensing in precision and variable-rate spraying typically combines light detection and ranging (LiDAR) or ultrasonic ranging for canopy geometry with wheel encoders and an inertial measurement unit (IMU) for motion estimation, and it may add cameras when plant detection or disease targeting is required. Common failure modes include wind and turbulence that deflect droplets and canopy occlusion that leads to under- or over-estimation of target structure. A key mechanistic bottleneck is that deposition depends on coupled airflow, platform motion, and droplet dynamics; therefore, improved target detection alone does not guarantee improved deposition unless actuation and airflow control compensate for wind and motion effects [91]. Cross-study comparability is also limited when evaluations omit critical parameters such as nozzle configuration, boom height, meteorological conditions, or standards-aligned measurement distances [90]. As shown in Table 5, representative spraying approaches differ in sensing and control structure, leading to distinct tradeoffs in dosing responsiveness, selectivity, drift behaviour, and evaluation metrics across agricultural settings. To improve engineering interpretability, Table 5 makes explicit the minimum descriptors needed to compare spraying systems across studies, namely, the sensing stack, the control logic, the implement/nozzle specification, and the principal outcome metrics with units. In particular, implement reporting is expanded to include nozzle type/configuration, pressure, flow rate, boom or air-assist geometry, and timing-related quantities because deposition outcomes depend on actuation and spray physics, not only on target detection [35,53,66,70,73].

4.4. Harvesting

Robotic harvesting requires detection, localisation, approach planning, and compliant detachment while preserving produce quality. Harvesting reviews emphasise that occlusion by foliage, variable illumination, and cluttered orchard backgrounds create persistent perception difficulty and propagate into grasp instability and collision risk [91]. Perception in robotic harvesting commonly relies on deep-learning-based detection or instance segmentation, often fused with red-green-blue depth (RGB-D) sensing to estimate target pose. Typical failure modes include missed detections under occlusion, false positives in cluttered scenes, and pose uncertainty that produces unstable grasp planning and execution. End-effectors span rigid grippers, suction tools, and hybrid designs; soft robotic grippers are investigated to increase compliance and reduce bruising, while hybrid suction–finger mechanisms aim to balance contact area, stability, and robustness to target variability [92,93].

Table 7. Representative harvesting end-effector approaches and reported evaluation characteristics.

Attribute	RGB-D + Suction	RGB-D + Soft Gripper	Hybrid Suction + Gripper
Typical sensors	RGB-D/stereo; optional wrist camera; vacuum feedback where used	RGB-D + tactile/force feedback where used	RGB-D/multi-view RGB-D; optional wrist camera; vacuum/force feedback
Typical model/control	Detection + 3D pose + approach/detach control	Segmentation/keypoints + grasp planning + compliant control	Detection + stem/keypoint localisation + trajectory planning + coordinated grasp/detach
Implement/tool	Suction cup/head (cup size, vacuum level, detachment method)	Soft/pneumatic gripper (finger count, stroke, compliance, grasp force)	Hybrid suction–gripper/cutter (geometry, DOF, compliance, actuation sequence)
Final detachment mechanism/device	Pull-off/traction detachment; or suction + cutter (knife/scissors/oscillating blade) when stems/pedicels require cutting	Twist-and-pull; pull-off; optional cutting tool (scissors/knife) depending on crop	Integrated cutter (scissors/knife/oscillating saw) or twisting module; coordinated suction stabilisation + cut/twist
Reported strengths	Reduced contact; some occlusion handling	Damage mitigation	Stability under clutter
Reported weaknesses	Wet/dust surfaces; leaf interference	Pose error sensitivity	Complexity; calibration
Evaluation setting	Greenhouse/orchard	Orchard	Orchard
Metrics used	Harvest success (%); cycle time (s fruit−1); damage (%); detachment success (%)	Harvest success (%); damage/slip; cycle time	Harvest success (%); cycle time; damage (%); retry rate; grasp/localisation error
Refs.	[94,98,99]	[92,95,97,100,101,102]	[52,92,93,96,103,104]

summarises representative harvesting end-effector approach categories and their reported evaluation characteristics, including sensing modalities, control strategies, implement types, operational strengths and weaknesses, and commonly used performance metrics. To improve engineering interpretability, Table 7 states more explicitly the sensing stack, the perception/planning or compliant-control strategy, the end-effector configuration, and the outcome metrics with units. This is important because harvesting performance is jointly determined by perception quality, detachment strategy, contact compliance, and damage-sensitive handling rather than by detection accuracy alone [52,94,95,96,97].

Localisation strategies vary with the environment. In orchards, systems may use the global navigation satellite system (GNSS) with real-time kinematic (RTK) corrections for coarse positioning, but they still depend on local sensing for branch avoidance and accurate base placement. In greenhouses, navigation more often relies on simultaneous localisation and mapping (SLAM) or fiducial markers. Planning must explicitly account for collision constraints and uncertainty: When targets are only partially visible, the robot may need active perception through repositioning or adopt conservative approach strategies to maintain safety. Evaluation typically reports picking success rates, cycle times, and produce damage, yet cross-study comparability is limited by differences in orchard structure and occlusion prevalence, as well as inconsistent definitions of a “successful” harvest [105]. A key mechanistic bottleneck is uncertainty propagation: because contact actions are largely irreversible, perception and pose errors can bruise produce or entangle branches, which encourages conservative planning and reduces throughput unless uncertainty-aware planning is paired with compliant control [106].

5. Enabling Technologies Across Tasks

5.1. Localisation and Mapping in Agriculture

Localisation in agricultural robotics is largely conditioned by the availability and accuracy of the global navigation satellite system (GNSS) and by environmental effects that degrade onboard sensing. A global reference frame is often provided by GNSS with real-time kinematic (RTK) corrections for mapping and repeat operations; however, in under-canopy and protected cultivation conditions, GNSS can be unreliable, and visual odometry, light detection and ranging (LiDAR)-based simultaneous localisation and mapping (SLAM), and multi-sensor fusion are therefore adopted. Reviews of LiDAR in agriculture emphasise its broad use for canopy sensing and environment perception while also underscoring failure conditions such as dust, aerosols, and partial occlusion that reduce returns and complicate scan matching [20]. GNSS-restricted navigation evaluations, including field robot competitions, further highlight the operational difficulty of relying only on relative sensing and state estimation, particularly under missing plants and irregular rows [9]. Robust localisation is therefore typically achieved through sensing redundancy and explicit failure handling, rather than by relying on a single estimator tuned to nominal conditions [20].

5.2. Perception and Learning

Across tasks, perception is increasingly implemented using deep learning for detection and segmentation because plant-scale discrimination is required for weeding, targeted spraying, and harvesting. However, robustness is often limited by domain shift. A recent review on transferability and robustness in proximal and unmanned aerial vehicle (UAV) crop imaging argues that such failures can be interpreted along shift axes such as scene, sensor, acquisition protocol, and time and that credible robustness claims require evaluation protocols that hold out entire sites and seasons rather than relying on random splits [18]. This shift-aware framing is supported when benchmark and task-specific datasets capture real-world heterogeneity. For example, Agriculture-Vision provides multi-band imagery with annotated field patterns for segmentation evaluation under operational variability [76]. Field-robot datasets spanning growth stages and revisit scenarios provide structured variation relevant to weeding and localisation, enabling assessment under occlusion and appearance change [85]. For harvesting, datasets collected under production-like conditions with clutter and occlusion have been introduced to reduce the gap created by limited representative annotations [107]. Overall, the evidence suggests that dataset design and evaluation splits are first-order determinants of deployment relevance.

5.3. Planning and Control in Unstructured, Deformable Environments

Planning and control in agricultural robotics are required to accommodate occlusion, deformable vegetation, and uncertainty in both the robot state and the target state. In navigation, row structure is often assumed for path generation, but gaps and irregularities necessitate state machine logic and reactive recovery behaviours; competition rules that penalise crop damage further indicate that tight-clearance control must be maintained under uncertainty, rather than focusing only on waypoint tracking [9]. In harvesting, planning is inherently performed under incomplete information because fruits are frequently partially occluded and branches create clutter, which motivates collision-aware planning and repeated replanning as the scene changes with plant motion. Reviews of fruit-picking robots similarly suggest that persistent occlusion and dynamic lighting are decisive in real deployments, implying that uncertainty and safety margins should be incorporated into planning [91]. Figure 4 synthesises the main environmental factors, failure modes, and mitigation strategies across the autonomy stack in agricultural robots by explicitly linking field conditions to representative failure mechanisms and the corresponding mitigation and validation practices. Comparable requirements are reported in greenhouse navigation, where confined spaces motivate hybrid schemes that couple global planning with local obstacle avoidance and where localisation quality governs feasibility of narrow-gait traversal [42,108].

5.4. Manipulation and End-Effectors

Manipulation in agricultural harvesting and selective treatment tasks is contact-rich and damage-sensitive. Soft robotics and compliance control are widely studied as mechanisms to tolerate pose uncertainty and reduce damage risk. Field-evaluated soft gripper studies in fruit harvesting report that compliant fingers and hybrid suction–finger designs can improve conformability and detachment robustness, but they also imply new control requirements: Compliance must be managed so that grasp forces remain within safe bounds while still achieving reliable detachment [92,93]. From a mechanical-design perspective, end-effector design is tightly coupled with planning, since increased compliance can expand the set of feasible approach trajectories but also increases uncertainty in contact states, thereby shifting performance requirements toward sensing and control during interaction [92].

From a mechanical-engineering perspective, the performance of agricultural end-effectors cannot be assessed adequately from task labels alone; the engineering reference value depends on the reporting of quantitative design descriptors. For soft harvesting grippers, relevant parameters include the constituent material and its compliance-related properties, actuation principle, finger geometry, characteristic opening/closing stroke, achievable contact or clamping force, contact pressure, payload-to-weight ratio, and response time, since these quantities govern fruit damage risk, grasp stability, and repeatability [109,110]. Recent reviews also note that soft grippers are still reported in a non-homogeneous manner, which makes direct comparison difficult and limits transferability across crops and operating conditions [109,110]. Similarly, for spray implements, engineering interpretation requires reporting the nozzle type/orifice specification, working pressure, per-nozzle flow rate, and nozzle-to-canopy assignment, because these parameters directly affect deposition uniformity, dosage control, and drift behaviour. As an example, the LiDAR-guided orchard sprayer in [53] employed flat-fan nozzles operated at 0.3 MPa, with a reported flow rate of 0.59 L min⁻¹ per nozzle. Accordingly, end-effectors and operating implements should be discussed not only by function but also through the key mechanical and fluid-delivery parameters that determine field performance and design reproducibility.

5.5. Safety, Human–Robot Interaction, Ergonomics, and Regulatory Constraints

Safety in agricultural robotics is shaped by the use of heavy machinery, sharp implements, chemical agents, and frequent operation in close proximity to workers within shared workspaces. Safety principles and verification and validation considerations for partially automated, semi-autonomous, and autonomous agricultural machinery are defined in the International Organization for Standardization (ISO) 18497 series, where autonomy claims are expected to be supported by risk assessment, hazard analysis, and validation evidence [22]. Functional safety for safety-related control systems in agricultural machinery is addressed in ISO 25119, which frames the design and development of safety-related parts of control systems [111]. For unmanned aerial vehicle (UAV) operations, aviation regulations further constrain feasible operating envelopes, and regulator guidance emphasises visual line-of-sight requirements and other routine operational limitations [36,112]. In the robotics literature, safety is commonly implemented through emergency stop functions, proximity sensing, speed limiting near humans, and explicit field protocols such as exclusion zones, yet safety validation is often reported less consistently than task performance [22].

Human–robot interaction in agriculture should be treated as a core autonomy function rather than only as a supervision interface. Recent reviews show that agricultural HRI spans teleoperation and supervisory control, collaboration/cooperation modes, wearable sensing, and vision-based non-verbal communication, while harvesting and spraying remain the most frequently studied human–robot collaboration applications [113,114]. In noisy, cluttered, or otherwise communication-constrained field environments, practical HRI mechanisms such as gesture-based commands and dynamic movement recognition are particularly relevant because they provide intuitive communication when speech interaction is unreliable. Recent field studies demonstrate real-time gesture and movement vocabularies for actions such as lock/unlock, follow, stop, return, and emergency stop, while maintaining explicit speed and human–robot separation constraints during operation [115,116].

Safety in human–robot interactive agricultural operations should also be analysed together with ergonomics rather than only through obstacle avoidance or emergency-stop logic. Review evidence shows that safety research on automated agricultural machinery spans environmental perception, risk assessment and mitigation, and human factors/ergonomics, while HRC roadmaps in agriculture identify safety, ergonomics, awareness, and productivity as co-equal design objectives [114,117]. In agricultural HRI, ergonomics matters because collaborative systems may reduce physical workload while still exposing workers to posture-related and musculoskeletal risks during harvesting, handling, or supervision tasks. Accordingly, evaluation should report not only task success but also workload, perceived safety, trust, operator role, training, and, where appropriate, posture- or fatigue-related measures [113,118]. Recent work on explainable AI-enhanced human activity recognition further suggests that transparent Human Activity Recognition (HAR) models can support safer and more trustworthy collaboration by making activity classification and sensor relevance more interpretable in field settings [119].

These gaps motivate the use of explicit safety, HRI, and ergonomics reporting criteria, together with evaluation ladders.

Table 8. Common sources of domain shift affecting agricultural robotic systems and typical mitigation and validation strategies. For interpretive clarity, the listed factors include both crop-related/biological variability and operational/environmental limitations.

Source of Shift	Symptom	Affected Module	Mitigation	Validation Method	Task(s) Impacted	Refs.
Seasonality	Appearance or structural changes across months	Perception, planning	Shift-aware training; temporal holdouts	Leave-one-season evaluation	M/W/H	[18,120]
Illumination	Shadows, saturation, exposure mismatch	Perception	Radiometric normalisation; data augmentation	Controlled illumination splits	M/W/H	[81,121]
Growth stage	Canopy closure; increased occlusion	Perception, navigation	Active perception; robust row models	Growth-stage holdout	M/W/H	[85,122,123]
Cultivar variability	Morphology and colour differences	Perception	Domain adaptation; fine-tuning	Leave-one-cultivar evaluation	M/W/H	[18,124]
Soil dust/aerosols	LiDAR return degradation	Localisation, perception	Sensor fusion; filtering	Dusty vs. clean comparative trials	M/W/S/H	[20,125,126]
Canopy occlusion	Target visibility loss	Planning, harvesting	Uncertainty-aware planning; repositioning	Occlusion-stratified tests	M/W/H	[127,128]
Weather (wind)	Plant motion; spray drift changes	Planning, spraying	Drift mitigation; speed adaptation	Standards-aligned drift tests	S	[90,129]

summarises common sources of domain shift in agricultural robotics, the operational symptoms they produce, the autonomy modules most affected, and representative mitigation and validation strategies reported in prior work. For clarity, the sources of shift summarized in Table 8 may be viewed broadly as either crop-related or biological variability, such as seasonality, growth stage, and cultivar differences, or as operational and environmental limitations, such as illumination changes, canopy occlusion, dust, and wind effects. Nevertheless, some factors remain cross-cutting across tasks and deployment settings. To avoid over-generalisation across task classes, Table 8 indicates the primary task(s) most directly impacted by each factor or failure mode, using M (monitoring), W (weeding), S (spraying), and H (harvesting). These tags are intended to indicate dominant relevance rather than strict exclusivity, since several factors remain cross-cutting across multiple agricultural robotic tasks.

Field deployment readiness depends not only on autonomy performance but also on labour context, ethical acceptability, and legal compliance. In practice, agricultural robots must be deployed with clear operator roles, supervision assumptions, training procedures, and accountability structures, rather than being assessed only through nominal task success [117,130]. Deployment is also shaped by whether the system reduces, redistributes, or increases human workload in realistic operations, and by whether the residual risk in shared human–robot workspaces is socially and operationally acceptable [113,130]. Accordingly, evidence of deployment readiness should include not only technical validation but also explicit reporting of supervision requirements, operator responsibility, safety procedures, and the relevant regulatory context. For the same reason, studies should distinguish clearly between lab prototypes, field prototypes, pre-commercial pilots, and commercially deployed systems, since nominal task success does not by itself establish deployment maturity or market readiness.

In conclusion, evaluation is a cross-cutting theme across perception, localisation, planning, control, tooling, safety, and HRI. For clarity and readability, these shared considerations are addressed in a dedicated discussion of evaluation metrics, practices, protocol considerations, benchmarking progression, and reporting guidance, as presented in Section 6.

6. Evaluation Practices and Benchmarking

Meaningful evaluation in agricultural robotics should be grounded in operational outcomes rather than perception accuracy alone. For monitoring, predictive accuracy should be complemented by measures of spatial registration, temporal consistency across repeated runs, and uncertainty estimates that directly affect downstream decisions. For weeding, performance should be reported in terms of weed suppression efficacy alongside crop injury rates and operational throughput, with crop spacing and weed density clearly specified. For spraying, evaluation should quantify deposition distributions and drift while reporting meteorological conditions, since these factors strongly influence outcomes. In this context, International Organization for Standardization (ISO) guidelines for spray distribution in tree and bush crops and for field measurement of spray drift provide measurement principles that can improve cross-study comparability when evaluation protocols are aligned with the standards [89,90]. Harvesting systems should report picking success, damage rates, and cycle time or throughput, with orchard or greenhouse conditions and occlusion prevalence clearly specified. Across tasks, reliability indicators such as uptime and failure-recovery frequency, as well as safety indicators such as near-miss events and human-proximity triggers, are also important for deployment relevance but remain inconsistently reported in the literature.

Table 9. Core common evaluation metrics used to assess agricultural field robots across tasks, together with their definitions, units, primary operation(s), and key measurement protocol notes. Abbreviations: M = monitoring; W = weeding; S = spraying; H = harvesting.

Metric Name	Definition	Unit	Primary Operation(s)	Measurement Protocol Notes
Spatial registration error	Difference between mapped and true location	m/cm	M/S	Report GNSS/RTK vs. SLAM; specify reference frame
Trait estimation error	Deviation from ground truth trait	Trait units	M	Specify sampling protocol and ground truth method
Weed suppression efficacy	Fraction of weeds removed/suppressed	%	W	Report weed density and growth stage
Crop injury rate	Fraction of crop plants damaged	%	W	Define “damage” and inspection timing
Deposition uniformity	Variability of deposit across canopy zones	Defined by protocol	S	Specify sampling locations; align with ISO if used
Drift quantity	Off-target drift at defined distances	Protocol-defined	S	Record weather; use ISO 22866 distances if applicable
Harvest success	Fraction of targets successfully harvested	%	H	Define “success” and failure categories
Damage proxy	Bruising/cut/quality downgrade rate	% or score	H	Define assessment method and delay
Throughput	Targets processed per unit time/area	Targets/h or ha/h	All tasks	Report downtime and intervention time
Intervention frequency	Number of operator interventions	Count/time	Autonomy	Define what constitutes an intervention

summarises the core evaluation metrics used to assess agricultural field robots across tasks, together with their definitions, units, task relevance, and key measurement protocol notes.

Fair comparison across studies requires that results be explicitly stratified by evaluation setting (e.g., laboratory, greenhouse, or open field) and by the environmental conditions under which data are acquired. Robustness-focused imaging reviews indicate that acquisition protocol choices and evaluation design can dominate apparent performance differences, and therefore, cross-study comparisons should not be treated as equivalent unless protocols and domains are aligned or the domain shift is explicitly modelled [18]. Accordingly, studies should report key operational design domain properties, including crop type, growth stage, illumination regime, canopy structure, and soil or terrain conditions, and generalized claims should be avoided when evidence is limited to single-site or single-season trials.

Bias in experimental evaluation can be reduced through cross-season, cross-field, and cross-cultivar testing, ideally using leave-one-site-season holdouts. Figure 5 illustrates the proposed benchmark ladder for agricultural robotics, showing the progression from component-level and laboratory testing to greenhouse, single-season field, and multi-season field evaluation with increasing system complexity and operational realism.

Dataset and benchmark studies demonstrate the value of longitudinal, multi-condition data collection. For example, field-robot datasets recorded across growth stages with calibrated sensors enable evaluation under changing plant morphology and occlusion [85], while multi-location monitoring datasets support shift-aware assessment across sites and seasons. In harvesting, benchmark datasets for densely fruited crops are designed to capture production-like variability and enable more comparable method evaluation [107]. Nevertheless, many robotics studies still rely on random train–test splits from a single field campaign, which mainly tests in-distribution performance and provides limited evidence of transferability [85]. A community reporting checklist should therefore require four elements:

• Autonomy-stack reporting should specify what is automated and what remains operator-assisted.
• Protocol reporting should document sensor calibration, data splits, and the evaluation environment.
• Safety reporting should summarise hazard-analysis elements, implemented safety mechanisms, and fail-safe behaviours.
• Reproducibility reporting should state dataset and code availability and provide key parameter settings.

Quantitative reporting should also standardize units and benchmark bases by expressing cost in currency per hour and, where relevant, per hectare; endurance in minutes or hours per charge/tank together with normalized energy/fuel intensity; and all relative comparisons with explicit statements of comparator, payload, working width, speed, and field conditions [9,22,29]. These requirements align with safety standards for autonomous agricultural machinery, where verification and validation principles are emphasised and safety claims are expected to be supported by explicit evidence and testing procedures rather than informal statements [22].

Table 10. Recommended reporting checklist for reproducible evaluation of agricultural field robots, outlining the key items to report, their rationale, suggested reporting practice, common pitfalls, and representative supporting references from the benchmark, standards, and methodological literature.

Item Authors Should Report	Why It Matters	How to Report	Common Pitfalls	Refs.
Crop, cultivar, growth stage	Domain shift and geometry	Describe and photograph; include dates	Vague crop description	[18,85,107]
Environment and weather	Affects perception and actuation	Record illumination, wind, and soil conditions	Missing meteorological context	[89,90]
Sensor suite and calibration	Reproducibility	List sensors; report calibration method	Omitting calibration details	[18,85]
Localisation method and failure handling	Safety and reliability	Describe estimator, fusion, and failover logic	Reporting only nominal accuracy	[9,22]
Actuation/tool details	Interpreting performance outcomes	Report nozzle/gripper/tool specifications	Missing tool configuration	[53,89,90]
Autonomy level and operator role	Interpreting autonomy claims	Map to Table 2 taxonomy; specify operator actions	Ambiguous use of “autonomous”	[9,22]
Evaluation protocol	Comparability across studies	Describe data splits, repetitions, and baselines	Random splits only	[18,85,107]
Safety mechanisms and protocols	Worker and crop safety	Report E-stop, proximity sensing, and field procedures	No safety validation evidence	[22,111]
Dataset/code availability	Reproducibility	Provide links and licences for artefacts	Unavailable artefacts	[85,107]

presents a recommended reporting checklist for reproducible evaluation of agricultural field robots, outlining the key items to report, their rationale, suggested reporting practice, and common pitfalls to avoid.

7. Synthesis of Tradeoffs and Design Principles

The purpose of this section is to synthesise design tradeoffs and to distil design principles that generalise across tasks and platforms. Figure 6 presents a qualitative radar chart comparing the conceptual tradeoffs among UGVs, UAVs, gantry systems, and greenhouse robots across key agricultural robotics performance dimensions. For tasks, agricultural robots operate under coupled constraints that give rise to recurring tradeoffs. A key tradeoff exists between throughput, precision, and crop damage risk. This is particularly evident in harvesting, where higher throughput often requires faster motion and reduced sensing time, thereby increasing the likelihood of collision and produce damage under occlusion.

In contrast, more conservative operations improve safety and precision, but at the cost of reduced throughput [105]. A similar coupling is observed in weeding, where mechanical or thermal actuation near crops demands high precision. Conservative decision thresholds reduce the risk of crop injury but may leave weeds untreated, whereas more aggressive thresholds can improve weed removal at the expense of increased crop injury risk. In spraying, an additional tradeoff arises between operational speed, drift, and deposition quality: increasing coverage speed or airflow can elevate drift risk unless deposition is actively controlled, which motivates standards-based evaluation of drift and spray distribution [89,90]. Monitoring presents a related tradeoff between coverage and calibration robustness, as high coverage may require less controlled acquisition conditions, thereby increasing protocol-induced variability and reducing comparability across sites and time [106].

A second tradeoff arises between autonomy, supervision burden, and reliability. Systems operating in restricted domains, such as fixed-structure platforms or highly structured greenhouses, can often achieve higher autonomy with fewer exceptions, whereas open-field systems face greater environmental variability and therefore more frequently require operator intervention. Safety standards further indicate that increased autonomy should be supported by corresponding validation and risk assessment; otherwise, deployment risk is increased [22]. A third tradeoff concerns energy, payload, and operating time, which is particularly important for unmanned ground vehicles (UGVs) carrying implements and for unmanned aerial vehicles (UAVs) limited by flight endurance. In both cases, energy constraints directly influence sensor selection and the range of feasible exploration and planning strategies.

Several design principles can be formulated from the cross-task synthesis. First, irreversible actuation requires uncertainty-aware decision-making. In weeding, spraying, and harvesting, actuation errors can result in crop damage or incorrect chemical application; accordingly, perception modules should provide uncertainty estimates, and planning should incorporate these estimates rather than relying only on point predictions [18]. Second, evaluation should be aligned with the operational design domain. Component-level metrics alone do not reliably predict field-level performance under domain shift, and evaluation should therefore include cross-site and cross-season protocols together with clear reporting of environmental conditions [18]. Third, safety functions should be treated as core autonomy modules rather than secondary additions. Safety standards emphasise verification and validation, which implies that emergency stop functions, proximity sensing, and fail-safe planning should be explicitly specified, tested, and reported alongside task performance [22]. Fourth, platform choice should be treated as a primary design variable rather than an implementation detail, because platform constraints determine the feasible sensing, actuation, and safety envelope and therefore shape the interpretation of task-centric comparisons. These principles are intended as testable claims, since each can be linked to measurable outcomes under controlled changes in uncertainty modelling, evaluation protocol, safety validation, or platform archetype.

In addition to perception, evaluation, and safety, robust agricultural robot design also depends on mechanical embodiment and field reliability. For intervention tasks, tool-level parameters such as gripper force/stroke and compliance, laser power and spot diameter, and nozzle pressure and flow characteristics influence precision, selectivity, and repeatability, while reliability indicators such as uptime, failure frequency, wear or clogging, calibration drift, and maintenance burden determine sustained field usefulness [22,53,86,109,110].

Table 11. Design principles for robust and comparable agricultural robotic systems.

Principle	Rationale	Design Implication	Example Task(s)	Refs.
Uncertainty-aware actuation	Irreversible errors harm crops and people	Propagate uncertainty into planning and safety gating	Weeding, spraying, harvesting	[22,131]
Shift-aware evaluation	Random splits overstate robustness	Cross-site/season holdouts and metadata reporting	Monitoring, weeding	[18,124]
Standards-aligned deposition/drift reporting	Comparability requires protocol consistency	Use ISO measurement principles when applicable	Spraying	[90,129]
Safety as first-class module	Autonomy increases hazard exposure	Validated E-stop/proximity sensing and field protocols	All tasks near humans	[22,132]
Platform-aware claims	Platform constraints shape feasible autonomy	Report platform limits and operational design domain	All tasks	[9,22]
Embodiment-aware tool/implement specification	Actuation outcomes depend on nozzle/gripper/tool mechanics; missing specs undermine comparability and engineering transfer	Report key implement parameters (e.g., force/stroke, laser power/spot size, nozzle pressure/flow)	Harvesting, weeding, spraying	[53,86,109,110]
Mechanical reliability, maintainability and serviceability	Field deployment failures are frequently mechanical (clogging, wear, misalignment) and drive downtime/MTBF	Evaluate uptime, failures, drift, wear, and maintenance	All intervention tasks	[9,22,133]

summarises design principles for robust and comparable agricultural robotic systems, linking each principle to its rationale, practical design implication, representative task relevance, and supporting references while explicitly including both autonomy-stack and mechanical-engineering criteria.

A decision framework for selecting platform and toolchain can be formulated by mapping crop type and farm scale to task requirements and the associated autonomy-stack constraints. For large-scale field monitoring, unmanned aerial vehicles (UAVs) equipped with multispectral sensing are generally well suited, with primary emphasis placed on calibration, geo-referencing, and evaluation under domain shift. For under-canopy monitoring and intervention tasks, such as weeding and targeted spraying, unmanned ground vehicles (UGVs) are often more appropriate, with design priorities centred on row navigation, perception robustness under shadow and occlusion, and operational safety. For high-repeatability phenotyping or controlled interventions in research plots and specialised facilities, gantry and fixed-structure systems offer advantages through controlled sensor pose and improved longitudinal consistency. In greenhouse harvesting and inspection, the main constraints arise from structured but confined environments, where global navigation satellite system (GNSS)-denied localisation, reliable navigation, and safety in close proximity to workers become central. Hybrid systems are most suitable when monitoring and intervention are separated across time or space, although their effectiveness depends on reliable cross-platform calibration and data integration [17].

8. Open Challenges and Research Agenda

The reviewed literature indicates that agricultural robotics has progressed beyond isolated proof-of-concept systems toward more capable and deployment-aware platforms. Nevertheless, important limitations remain, including weak cross-site validation; heterogeneous reporting practices; incomplete mechanical specification; and uneven treatment of safety, HRI, and deployment readiness. In light of these gaps, the following research agenda summarises the main remaining technical and deployment-related challenges across the near-, mid-, and long-term horizons.

8.1. Near-Term Challenges

A near-term challenge lies in ensuring evaluation validity under domain shift, as many studies still rely on within-site random splits that do not assess transfer across seasons, sites, or cultivars. Robustness-focused reviews have argued that shift-aware evaluation is essential and that acquisition protocols and metadata can influence apparent performance gains as strongly as, or more strongly than, model choice. This suggests that reproducibility and protocol transparency are necessary conditions for meaningful progress [18]. A measurable advance would be the routine adoption of leave-one-site and leave-one-season evaluation for perception modules in monitoring and weeding, accompanied by reporting checklists that document acquisition settings and environmental context. Such evaluations could be supported through multi-site data collection under fixed protocols, pre-registered evaluation splits, explicit reporting of cross-site performance differences, and uncertainty-aware performance analysis.

A second near-term challenge is the integration of safety reporting into mainstream evaluation, as safety engineering is still often treated as a compliance requirement rather than as a measurable component of autonomy. Safety standards for autonomous agricultural machinery emphasise verification and validation, which indicates that safety mechanisms should be systematically tested and reported rather than described only in qualitative terms [22]. A measurable advance would be the adoption of field safety protocols that quantify emergency stop response, proximity sensing performance, and false stop rates under defined test conditions. Such evaluation could be implemented through controlled field trials with scripted human–robot proximity scenarios and independent measurement of stop distances and response times, reported alongside task performance metrics.

8.2. Mid-Term Challenges

A mid-term challenge is the achievement of long-term autonomy with maintainable calibration, as sensors drift over time, environments evolve, and prolonged field operation imposes sustained stress on both hardware and software. Longitudinal navigation studies and deployment-oriented investigations indicate that reliable transfer from simulation to field use requires systematic alignment and repeated evaluation; however, extending this reliability across multi-season operation further requires continuous calibration checks and effective fault detection [134]. A measurable advance would be the development of autonomy stacks capable of maintaining localisation and perception performance over multi-week deployments while providing explicit self-check mechanisms and defined recovery modes. This could be evaluated through multi-week field trials with periodic validation checkpoints, logged failures, and controlled perturbations, such as dust exposure and lighting variation, to assess recovery behaviour.

A second mid-term challenge is uncertainty-aware planning under occlusion and deformability, particularly in harvesting and close-proximity weeding. This challenge persists because occlusion creates partial observability, while plant motion can invalidate precomputed trajectories during execution. Reviews of harvesting systems identify occlusion and variable illumination as dominant barriers to reliable field performance, suggesting that meaningful progress will require planners that explicitly incorporate uncertainty and active perception rather than relying only on improved detection models [106]. A measurable advance would be the demonstration of harvesting systems that maintain stable performance across different levels of occlusion while incorporating explicit active perception strategies and failure recovery mechanisms. Such progress could be evaluated through occlusion-stratified field trials in unmodified orchards or greenhouses, with controlled occlusion measurement and systematic logging of failure categories.

8.3. Long-Term Challenges

A long-term challenge is the achievement of multi-season, multi-site field autonomy with rigorous safety assurance. This challenge persists because real deployment requires robustness to evolving environments, reliable remote supervision, and validated safety behaviour under rare but potentially severe events. Whole-cycle autonomous farming initiatives have demonstrated that autonomous operation can be achieved by combining off-the-shelf machinery with autonomy components in restricted domains; however, peer-reviewed analyses indicate that these successes depended on simplification of the environment and operating conditions, which suggests that broader generalisation remains unresolved [135]. A measurable advance would be the demonstration of autonomous operation across multiple fields and seasons with formalised safety cases and documented reliability, without requiring environmental simplification beyond standard agronomic practice. Such progress could be evaluated through multi-site living-lab deployments with gradual expansion of the operational design domain, supported by documented safety-case artefacts and independent auditing of incidents and near-miss events.

Another long-term challenge is the development of scalable benchmarking infrastructure that enables comparable and standards-aligned evaluation across agricultural robotic tasks. This challenge persists because field experiments are costly and highly context-dependent, while existing datasets and benchmarks often differ in crop type, sensing modality, and evaluation protocol. A measurable advance would be the broad adoption of benchmark ladders and shared test procedures, including standards-aligned protocols for spraying drift and deposition, as well as human-proximity safety testing, thereby enabling meaningful meta-analysis across studies. Such progress could be pursued through multi-institution benchmark campaigns that combine shared datasets with shared field protocols and cross-season repetitions. Figure 7 summarises the proposed roadmap for agricultural field robotics, outlining near-, mid-, and long-term milestones and their progression toward more robust, scalable, and credible autonomy. The roadmap is intended as a literature-grounded visual synthesis rather than a purely conceptual sketch, and the milestones are framed in terms of measurable evidence artefacts identified across the reviewed work, including shift-aware evaluation records, protocol-compliance artefacts, standards-aligned validation outputs, long-duration reliability logs, and formal safety-case documentation [18,19,22,85,89,90,106,107,134,135].

9. Conclusions

This review examined agricultural robotics through four principal tasks, monitoring, weeding, spraying, and harvesting, using complementary platform-based and autonomy-stack-based perspectives. The platform type, including unmanned ground vehicles (UGVs), unmanned aerial vehicles (UAVs), gantry and fixed-structure systems, greenhouse robots, and hybrid systems, was shown to constrain feasible sensing, actuation, and safety envelopes, therefore shaping how autonomy claims should be interpreted. In parallel, mapping each task from sensing to action onto the autonomy stack, including perception, localisation, planning, control, actuation, safety, and human-robot interaction (HRI), enables a more mechanistic comparison of failure modes, bottlenecks, and design choices across application domains.

Across these tasks, several recurring design tradeoffs were identified. In weeding and harvesting, throughput, precision, and crop damage risk are tightly coupled because actions are local and often irreversible. In spraying, drift and deposition quality emerge as additional coupled outcomes, which makes standards-aligned measurement essential for meaningful comparison. The autonomy level was also found to be closely linked to supervision burden and reliability, and higher autonomy without validated safety mechanisms increases deployment risk. Taken together, these patterns support design principles that emphasize uncertainty-aware actuation, shift-aware evaluation, safety as a core autonomy function, and platform-aware interpretation of results.

Compared with industrial robotics, agricultural field robotics must operate in only partially structured and seasonally changing environments, where deformable biological targets, occlusion, illumination variability, GNSS degradation, and terrain-induced slip, sinkage, and vibration alter both autonomy design and the meaning of reported performance. These conditions make deployment readiness depend not only on component accuracy but also on robustness to domain shift, tool-specific actuation quality, supervision burden, and safety behaviour in realistic field operations. For this reason, the present review places particular emphasis on operational design domains, implement/tool specification, and evaluation protocols rather than on perception or control accuracy alone.

A further synthesis emerging from the reviewed literature is that agricultural robotic systems should be interpreted according to deployment maturity rather than only nominal functionality. In practice, the evidence base differs substantially between lab prototypes, field prototypes, pre-commercial pilots, and commercial systems. The review therefore argues that future studies should report deployment maturity explicitly together with the scope of field evidence, including the number of sites and seasons, operating hours, failure categories, supervision requirements, and safety procedures. Without such evidence, strong component-level results should not be interpreted as proof of commercial readiness or broad field robustness. In this context, laboratory concepts are generally supported by controlled-environment or component-level validation, whereas field prototypes are typically evaluated through task-specific trials with limited site or seasonal coverage. Pre-commercial pilots provide more integrated demonstrations under realistic operational conditions. Commercial readiness, however, should be interpreted more cautiously and should ideally be supported by sustained multi-site or multi-season evidence, explicit reliability reporting, clear supervision requirements, and safety-validated deployment procedures rather than task success alone.

A central bottleneck to progress remains evaluation practice. Many studies continue to report strong component-level results under controlled conditions, while providing limited evidence of transfer across sites, seasons, or cultivars. Robustness-focused imaging studies indicate that credible deployment claims require evaluation protocols with spatial and temporal holdouts, together with clear reporting of acquisition settings and metadata. At the same time, safety standards for autonomous agricultural machinery make clear that validation should address hazards, verification procedures, and evidence of safe operation, rather than task success alone. For this reason, the proposed benchmark ladder and reporting checklist are presented as practical steps toward greater comparability and stronger deployment relevance.

Compared with earlier reviews that are often organised primarily by platform, sensing modality, or individual application area, the main contribution of this review is the joint synthesis of agricultural robotics through task, platform, autonomy-stack, and evaluation-setting lenses. This integrated framing makes it possible to compare systems not only by what they do but also by how they sense, act, fail, and are validated under different deployment conditions. The most important challenges that remain are shift-aware generalisation across sites and seasons, reliable localisation and planning under occlusion and deformability, standardised reporting of tool and safety performance, long-duration reliability with explicit fault handling, and clearer evidence of maturity and deployment readiness in real agricultural practice.

Overall, the reviewed advances indicate that agricultural robotics can contribute not only to higher productivity and more sustainable food production, consistent with SDG 2, but also to more efficient resource use and reduced waste, consistent with SDG 12. Moreover, the transition toward uncertainty-aware, resource-efficient, and electrified robotic systems strengthens the relevance of this field to SDG 13, particularly in the broader context of net-zero agriculture.

A realistic path toward robust field autonomy is likely to be progressive rather than abrupt. In the near term, deployment relevance can be improved through shift-aware evaluation, protocol transparency, and systematic safety reporting. In the mid-term, dependable long-duration autonomy will require maintainable calibration, fault detection, and uncertainty-aware planning under occlusion and deformability. In the long term, multi-season and multi-site operation will depend on scalable benchmarking infrastructure, formal safety cases, standards-aligned validation, and sustained field testing. Progress along this path would strengthen the role of robotics in precision agriculture by enabling more repeatable monitoring, safer and more selective intervention, and higher quality harvesting under realistic field variability.