MDPI Computation:从研究到部署——自主农业机械路径规划综述及可部署性评估框架(原标题:From Research to Deployment in Autonomous Agricultural Machinery: A Review of Path-Planning Technologies Against a Deployability Assessment Framework)

Logical Operator Operator

Search Text

Search Type

add_circle_outline

remove_circle_outline

Article Menu

Font Type:

Arial Georgia Verdana

Font Size:

Aa Aa Aa

Line Spacing:

  

Column Width:

  

Background:

Open Access Review

by
Sam Wane

1,
Redmond R. Shamshiri

1,2,
Wei Guo

3,
Haibo Chen

4 and
Fernando Auat Cheein

1,*

1

Engineering Department, Harper Adams University, Edgmond, Shropshire, Newport TF10 8NB, UK

2

Maersk Mc-Kinney Møller Institute, University of Southern Denmark (SDU), 5230 Odense, Denmark

3

College of Artificial Intelligence, Henan Agricultural University, No. 218 Ping’an Avenue, Zhengdong New District, Zhengzhou 450046, China

4

National Center for International Collaboration Research on Precision Agricultural Aviation Pesticide Spraying Technology, South China Agricultural University, Guangzhou 510642, China

Author to whom correspondence should be addressed.

Submission received: 29 June 2026 / Revised: 31 July 2026 / Accepted: 11 August 2026 / Published: 21 August 2026

Abstract

Global labour shortages in the agricultural sector, combined with diminishing arable land and a growing population, are driving investment in autonomous agricultural machinery. Autonomous systems that can navigate crop environments and perform planting, treatment, and harvesting alongside humans are required, but the gap between published research and commercially deployed systems remains wide across most operational scenarios. Why are agricultural robots still not widely deployed in real farms despite decades of research in autonomous navigation and path planning, and what is preventing full farm autonomy? This paper reviews the principal enabling technologies for autonomous agricultural integration, with a specific focus on path planning as the differentiator between research-stage and deployed systems. Current research in human–robot integration, open-field navigation, row identification and following, crop sensing, and power efficiency is synthesised and evaluated against a deployability criterion. A Deployability Assessment Framework is introduced, comprising structured tables that assign Technology Readiness Levels to twelve path-planning families and benchmark eleven commercial and research platforms against field-validated accuracy data. The analysis shows that point-to-point GNSS navigation has reached TRL 9 with over one million commercial units deployed, vision-based crop row following is at TRL 5–7 depending on crop and season, and whole-farm autonomy with dynamic re-planning is at TRL 3–5. The primary barriers are the absence of standardised evaluation benchmarks, the failure of perception models to generalise across seasons and crop types, and the decoupling of terrain and slip feedback from global path planners. Our review reveals that open-field GNSS navigation is commercially mature, but true whole-farm agricultural autonomy remains unsolved because current systems are not robust enough across seasons, terrain, sensing conditions, and operational transitions.

1. Introduction

In the West, farm ownership has become progressively consolidated under fewer, larger operators [1], while globally a structural agricultural labour shortage persists due to rural-to-urban migration, the unreliability of seasonal workforces, and the perceived hardship of farm work. These pressures create a direct demand for automation that can extend productive capacity without proportionally increasing headcount. Larger field sizes, tighter weather windows, and the agronomic requirement for precise, timely operations further accelerate this demand.

The complexity of the agricultural environment—unstructured terrain, seasonal change, unpredictable obstacles, variable soil conditions, and the co-presence of humans—explains the persistently low uptake of truly autonomous navigation for harvesting and crop treatment [2]. Reviews of navigation technology across plantation, livestock, and aquatic production confirm that no single platform or sensing modality covers all application types [2].

Path planning is the technological linchpin connecting a high-level farm management instruction (“spray field B”) to the low-level motor commands that safely execute it. It determines how efficiently a machine covers a field, how reliably it follows crop rows, how it handles headlands and obstacles, and, ultimately, whether a system is suitable for commercial deployment. Despite hundreds of published path-planning algorithms, the gap between laboratory demonstrations and commercially viable, field-deployed systems remains wide [3,4]. This review argues that deployability—the measurable readiness of a complete system for autonomous field operation under realistic conditions—is the correct lens through which to evaluate progress in agricultural robotics.

This paper makes three specific contributions: (1) a critical review of enabling technologies—HRI, open-field navigation, row following, crop identification, and power management—evaluated against their contribution to deployability; (2) a Deployability Assessment Framework comprising structured tables that assign TRL values to twelve path-planning families and benchmark eleven commercial and research platforms against field-validated accuracy data; and (3) an identification of four technical gaps—standardised benchmarks, seasonal perception generalisation, terrain-integrated planning, and regulatory pathways—that currently block the transition from TRL 6–7 to TRL 9 for row following systems.

The review is organised as follows. Section 2 covers the principal technology areas (Section 2.2, Section 2.3, Section 2.4, Section 2.5 and Section 2.6). Section 3 presents the deployability analysis and tables. Section 4 identifies outstanding technical and regulatory challenges, and Section 5 provides conclusions.

2. Analysis of Current Problems and State of Solutions

2.1. Review Methodology

This review was developed through a targeted literature search with a goal of approximately 100 journal papers focused on path planning and its derivatives in agricultural settings. Scopus and Web of Science were the primary bibliographic databases. A total of 88 papers were investigated once duplicates were removed, with the remaining references referring to company websites or ISO specifications. Grey literature, vendor white papers and blogs were read for orientation but were not cited. Some of the figures were generated or enhanced with the aid of AI. Selection criteria: journal articles in English, published from 2019 to 2026, in Agricultural Sciences, Engineering, and Computer Science. Articles exclusively on teleoperation or without relevance to in-field path execution were excluded; earlier foundational studies were also included where the field cites them by convention. The search string combined the following:

(title:("farming" OR "precision agriculture") AND abstract:("robot" OR "automated")) AND keywords:("crop row following" OR "path planning" OR "autonomous navigation" OR "human-robot interaction")

The four main resulting pillars of deployability: human–robot interface (HRI), navigation and autonomous movement, crop identification and handling, and power efficiency and environmental impact—are illustrated in Figure 1 and structure the review.

Figure 1. Summary of key areas: (1) HRI, (2) navigation and autonomous movement, (3) crop identification and handling, (4) power efficiency and environmental impact.

2.2. Integration with Humans and Existing Machinery (HRI)

The transition to field autonomy is constrained by the cost of replacing existing machinery and the time required for farm workers to adopt new operational practices. The scope of this transition across plantation, livestock, and aquatic applications is reviewed in [2]. Overall, HRI research accounted for less than 12% of the research papers reviewed, with the majority originating in the USA, Northern Europe, and Japan. Integration of the Robot Operating System (ROS) with legacy farm machinery is demonstrated in [5], which presents an effective graphical user interface enabling high-level commands (go to field, follow leader, go to charging station) across a mixed fleet of conventional and robotic vehicles. A modular hardware-software framework using ROS, standardised sensors, and replaceable 3D-printed chassis components has been validated across cotton-crop field trials [6].

Machines operating near humans must approach from predictable directions, maintain defined safety distances, and respond to operator gestures. A Human-Aware Module (HAM) applied to soft-fruit picking [7] shown in Figure 2, models the human movement envelope using cameras to infer intent through gesture recognition, and legs detected by the LiDAR coverage as shown in (a) blue and pink with the red arrow in (b) demonstrates the leg’s detection and likely movement direction. The new generation of collaborative robots (cobots [8]) and sensitive electronic “skin” [9] (Airskin [7]) are making close-proximity human–robot operation viable on agricultural manipulators. Risk assessment frameworks for autonomous agricultural machinery are becoming standardised practice [10], and deep-learning-based human-intention recognition with safety zones is advancing rapidly [11]. A comprehensive review of human–robot collaboration (HRC) in agriculture covering 55 studies identifies sensing infrastructure, safety certification, and operator training as the dominant barriers to deployment [12].

Figure 2. Human-aware navigation module for cooperative fruit-picking (a) Detection coverage areas according to the sensor used. (b) Inputs and outputs of the camera-based skeleton detector (blue) and the LiDAR-based leg detector (red) [7].

Fleet coordination extends HRI beyond individual platforms. A decentralised framework for allocating vehicles, implements, and raw materials dynamically across a farm, accounting for soil stability, time horizon, and shared ownership, is proposed in [13]. The leader–follower architecture in [14] allows a human-driven tractor to lead a swarm of autonomous robots, with the swarm pattern, spacing, and headland-turning logic managed autonomously. Route descriptions using controlled natural language and ISO 8373:2021 vocabulary [15] are encoded in the Mission Planner [16], enabling routes from the machine shed through gates and across fields to be expressed in a human-readable farm description.

The analysis of trailer dynamics during headland turning [17] addresses a common cause of crop damage and wasted headland in towed implements, providing analytical paths for N-trailer vehicles that prevent corner cutting. Vehicle rollover risk as a function of terrain slope, speed, and turning radius is tabulated in [18] and minimised through dynamic path constraints in [19].

2.3. Open-Field Navigation

Open-field navigation—locomotion without crop row constraints—is the most mature segment of agricultural autonomy, covering 27% of the reviewed papers, with a majority of them originating from China and Europe. Summarised in Figure 3, navigation relies on Global Navigation Satellite Systems (GNSS) to derive position and heading, with the pure-pursuit algorithm [20] the most widely deployed controller. A comprehensive review of path-planning methods (pure pursuit, Linear Quadratic Regulator (LQR), Fuzzy Logic, Model Predictive Control (MPC)) and motion-control methods (Proportional–Integral–Derivative (PID), neural networks, sliding mode) in an agricultural context is provided in [21].

Figure 3. Path-planning and motion-control taxonomies for unmanned agricultural tractors [21].

The heading accuracy achievable from two spatially separated GNSS antennas fused with an Inertial Measurement Unit (IMU) is critically assessed in [22], which proposes an adaptive Kalman filter that significantly reduces bias on narrow, bumpy terrain. Centimetre-level accuracy using the Quasi-Zenith Satellite System correction signal has been validated in mountainous farmland [23]. RTK-GNSS evaluation on dynamic and static platforms against a total station ground truth is described in [24].

A complete navigation stack (“FiWare”/Smart Navigation Manager) is developed in [25,26]: a high-level GOTO-GOAL command produces a motion profile, including orientation; Dubins and Spiral controllers manage Ackermann steering; and LiDAR-segmented safety zones trigger real-time path re-planning. The Reeds–Shepp curve [27] improves coverage efficiency over irregular field boundaries. Smooth headland turns using Dubins curves and spiral controllers with actuation-rate constraints are formalised in [28]. The entry/exit path planner in [29] uses the A* algorithm with vehicle turning-radius constraints and was validated on a 60 kW tractor.

Complete Coverage Path Planning (CCPP) has received substantial recent attention. The work in [3] presents a systematic CCPP approach for wheeled robots addressing field entry, irregular boundaries, and headland coverage; Ref. [30] extends this to 3D fields with terrain-inclination effects on soil erosion and energy consumption. The review in [4] proposes a three-stage methodology: guidance-track generation, traversal sequencing, and smooth drivable path generation. Headland coverage planning with continuous curvature and full tractor-implement geometry constraints is solved in [31], the first work to guarantee that field borders are not crossed during manoeuvres.

Obstacle avoidance using binarised Google Maps imagery is demonstrated in [32,33]; the minimum-snap algorithm adds vehicle kinematics (minimum curvature, jerk, and snap constraints) to obstacle-circumnavigation planning [34]. A hybrid semantic–topological–metric map with a novel switching algorithm [35] enables low-energy path planning by incorporating surface type and inclination. Multi-robot cooperative planning, where a UAV and UGV coordinate to locate and service insect traps in olive groves using YOLO-based identification and fiducial-marker landing, represents an emerging frontier [36]. It must be mentioned that the use of UAV imagery, proximal sensing, multispectral or hyperspectral data, and AI-based crop monitoring can provide useful information for autonomous robots, such as disease distribution, weed patches, crop growth status, and prescription maps, giving a dual use for real-time UAV imagery.

Path planning in agriculture is often task-dependent. For example, the best path for crop scouting, disease monitoring, selective spraying, weeding, or harvesting may not be the same. Table 1 uses information in [37] to show the relationship of different agricultural tasks with different navigation requirements, such as coverage efficiency, row following accuracy, target localisation, repeatability, and avoidance of crop damage.

Table 1. Agricultural tasks and corresponding navigation requirements.

A recent review of the control systems of agricultural robotics [38] emphasises that navigation requirements vary significantly across field operations. Tasks such as scouting prioritise coverage efficiency and mapping