Advances in Perception, Autonomous Operation, and Collaborative Systems for Smart Orchard Robots

1. Introduction

Orchard production is one of the most labor-intensive components of agricultural production systems and involves diverse operations, including blossom pollination, pest and disease control, weeding, inspection, harvesting, and transportation. Compared with field-crop production, orchard operations are highly seasonal, have narrow operational windows, involve spatially dispersed targets, and remain heavily dependent on manual labor. In recent years, rising labor costs and a declining agricultural workforce have placed substantial pressure on conventional orchard-management practices. In particular, reliance on manual labor alone may be increasingly insufficient to meet the requirements of large-scale, refined, and highly efficient production in harvesting, spraying, pollination, and weeding. Accordingly, the use of robots, sensors, autonomous navigation, machine vision, and intelligent optimization algorithms to increase orchard automation has become a major direction in the development of smart orchards [1,2,3].

With the development of smart and digital agriculture, agricultural robots are evolving from single-purpose mechanized equipment into intelligent systems capable of environmental perception, autonomous mobility, precision execution, and task-level decision-making. Previous studies have indicated that autonomous driving systems, intelligent agricultural robots, and multi-robot systems are increasingly being applied to conventionally manual tasks such as seeding, spraying, and harvesting, while agricultural production is progressively moving toward unmanned and intelligent operation [4]. In orchard environments in particular, robotic technologies have been used for fruit harvesting, flower pollination, variable-rate spraying, intra-row weeding, environmental monitoring, and yield estimation, thereby offering technical pathways to alleviate labor shortages, improve operational efficiency, and reduce resource consumption [5,6].

Existing studies have demonstrated a broad range of orchard-robot applications. Harvesting robots mainly address fruit recognition, three-dimensional localization, manipulator path planning, and end-effector control [7]. Spraying robots emphasize target recognition, variable-rate pesticide application, and improved utilization of spray liquid. Pollination robots focus on flower detection, identification of suitable pollination windows, and precise pollen delivery. Weeding robots primarily address weed recognition between rows or individual trees, path tracking, and operational safety [8,9]. For example, research on spraying robots for jujube orchards has noted the large scale of China’s jujube industry and the important role of high-density dwarf jujube orchards in Xinjiang. Mid-season management in these orchards still depends heavily on manual labor and is affected by pronounced seasonal and structural labor shortages, creating a clear practical demand for automated orchard robots [10].

At the same time, the capabilities of a single robot the capabilities of a single robot are no longer sufficient to meet the production requirements of large orchards. In extensive orchards, densely distributed task locations, or geographically distributed orchard systems, a single robot often suffers from low operational efficiency, limited coverage, and long task-completion times. Multi-robot collaboration, robot-UAV collaboration, and cross-orchard task scheduling have therefore emerged as new research directions in smart orchards [11,12]. Related studies have shown that agricultural robots and UAVs can be applied to pollination, harvesting, spraying, and monitoring. Ground robots are well suited to close-range precision operations, whereas UAVs are advantageous for aerial spraying and inspection. Coordinated use of the two platforms can exploit their complementary strengths and improve the efficiency of complex agricultural operations [13,14,15].

Consequently, research on intelligent orchard robots is no longer confined to a single mechanical structure or a single-purpose machine. Instead, it is increasingly oriented toward an integrated technological system comprising multimodal perception, autonomous navigation, precision operation execution, task scheduling, and intelligent decision-making. Achieving robust perception, reliable mobility, accurate execution, and efficient collaboration in complex orchard environments has become a key requirement for advancing orchard robots from experimental validation to practical deployment [16,17,18].

Orchard environments combine semi-structured and unstructured characteristics. Fruit trees are generally planted in rows and therefore exhibit a degree of spatial regularity; however, canopy morphology, foliage density, fruit distribution, terrain undulation, weed cover, and working-lane width remain highly variable. These environmental characteristics create substantial challenges for orchard robots in perception, localization, navigation, and operation execution [19]. In contrast to indoor or standardized industrial environments, orchard robots must not only identify operation targets but also understand tree structure, traversable areas, obstacle locations, and the spatial relationships between operation targets and robotic actuators.

First, complex orchard environments impose major challenges on perception systems. Fruits, flowers, branches, trunks, leaves, and weeds vary considerably in color, morphology, and spatial distribution and are frequently affected by occlusion, overlap, illumination variation, and background interference. For harvesting robots, fruit-localization errors directly affect grasping success and fruit-damage rates. For pollination robots, the accuracy of flower recognition and pollination-position control influences pollination performance. For weeding and spraying robots, accurate recognition of trunks, weeds, canopies, and target plants is essential for operational safety and resource-use efficiency [20,21]. Studies on semantic mapping have shown that conventional simultaneous localization and mapping (SLAM) can provide relatively accurate geometric maps, but geometric information alone is insufficient to support autonomous orchard operations. Robots must also understand semantic information, such as trees, terrain, and traversable regions, within the map. Second, orchard localization and navigation are challenged by unstable global navigation satellite system (GNSS) signals, sensor degradation, and dynamic environmental changes [22]. GNSS performs well in open farmland, but canopy occlusion, multipath interference, and terrain shielding in orchards can reduce positioning accuracy or cause signal loss. Inertial navigation can provide continuous short-term pose estimates but accumulates drift. Vision-based navigation provides rich environmental information but is vulnerable to illumination, weather, and motion blur. Two-dimensional LiDAR is relatively insensitive to lighting but captures only planar information and may provide insufficient geometry in complex canopies and undulating terrain [23]. To address these issues, current studies increasingly employ multi-sensor fusion involving LiDAR, RGB-D cameras, inertial measurement units (IMUs), wheel odometry, and adaptive Monte Carlo localization (AMCL) to improve localization stability and path-tracking performance in GNSS-constrained orchard environments.

Orchard operations also impose stringent requirements on navigation and execution accuracy. Spraying, pollination, weeding, and harvesting cannot be completed merely by reaching a target area; rather, robots must operate stably along specified paths and at prescribed distances and attitudes [24]. For example, a spraying robot in a high-density jujube orchard must follow predefined operation paths with high-precision cruising and point-specific stopping. Studies have implemented spray-route navigation using LiDAR SLAM, IMU-based inertial navigation, extended Kalman filtering (EKF), AMCL localization, A* path planning, and dynamic window approach (DWA)-based local obstacle avoidance [25]. Field experiments reported navigation-error indices for straight-line travel, point stopping, turning, and S-shaped path tracking, demonstrating explicit requirements for lateral-deviation and heading-angle control in practical orchard operation. Meanwhile, orchard-robot tasks are expanding from single-robot, single-task operation toward multi-robot, multi-task, and multi-scenario systems. Although multi-robot systems can improve operational efficiency, they introduce complex challenges involving task allocation, path conflicts, energy constraints, load balancing, and operation-time limits. Existing studies have emphasized the inherent coupling between task allocation and path planning in agricultural multi-robot operations. If task allocation is considered without accounting for obstacles, terrain constraints, and dynamic interactions among robots, the resulting solution may be difficult to execute effectively in real environments [26]. Joint optimization of task allocation and path planning is therefore an important direction for improving the feasibility and robustness of collaborative agricultural-robot operations.

Collaborative scheduling becomes even more complex in distributed or multi-orchard scenarios. Orchards may be geographically dispersed and differ in area, task density, and environmental conditions. In addition, robot and UAV operations may be subject to precedence constraints. For example, in the task-allocation problem involving pollination robots and spraying UAVs across multiple orchards, robots must complete pollination before UAV spraying begins, and this precedence requirement markedly increases scheduling complexity [27,28]. The simultaneous involvement of multiple agents, multiple orchards, heterogeneous task densities, and heterogeneous platforms further strengthens the combinatorial-optimization nature of these problems.

Overall, autonomous operation in complex orchard environments faces four principal challenges. First, perceptual targets are complex: fruits, flowers, branches, trunks, and weeds are readily affected by occlusion and environmental variation. Second, the navigation environment is difficult: GNSS constraints, uneven terrain, and canopy occlusion reduce localization stability. Third, operation execution requires high precision: robots must perform accurate, efficient, safe, and minimally damaging operations within confined spaces. Fourth, multi-robot collaborative scheduling is highly complex and must simultaneously account for task duration, path cost, load balance, energy constraints, and precedence relationships. Together, these challenges require orchard-robot research to progress from isolated module optimization toward system-level co-design.

To address the requirements of autonomous agricultural-robot operation in complex orchard environments, this review is organized around the technical chain of environmental perception, autonomous navigation, precision execution, and collaborative decision-making. In contrast to reviews focused on a single robot category or a single operation technology, this article emphasizes the technical interrelationships among the perception, mobility, execution, and decision layers and seeks to capture the overall trend toward systematic development of intelligent orchard robots.

The review first summarizes environmental-perception and semantic-understanding technologies for orchard robots. This section examines methods for recognizing and localizing representative objects, including fruits, flowers, branches, trunks, weeds, and obstacles, with particular emphasis on applications of machine vision, RGB-D sensing, stereo vision, LiDAR, point-cloud processing, instance segmentation, semantic segmentation, and VLMs in complex orchard scenes. Existing studies show that geometric maps alone cannot satisfy the requirements of autonomous operations. Semantic maps provide high-level information on tree layouts, target instances, and traversable regions and thus constitute an important foundation for autonomous navigation and task execution.

Second, this review summarizes localization, navigation, and scene-mapping methods for orchard robots. With emphasis on autonomous mobility in GNSS-constrained environments, this section discusses visual navigation, LiDAR navigation, SLAM-based mapping, multi-sensor fusion localization, semantic map construction, path planning, and path-tracking control. Related studies indicate that multi-sensor fusion can partially compensate for the limitations of individual sensors and improve navigation reliability in complex, occluded, and unstructured orchard scenes.

Third, this review summarizes precision-execution technologies for representative orchard operations, including harvesting, spraying, pollination, and weeding. Key issues include operation-target localization, manipulator control, variable-rate spraying, liquid pollination, intra-row weeding, wind compensation, error calibration, and system evaluation. Comparison across tasks shows that precision execution depends not only on actuator performance but also on close coordination among front-end perception accuracy, navigation stability, and operation-control strategies. Finally, multi-robot collaborative scheduling and intelligent decision-making are reviewed, with emphasis on multi-robot task allocation, joint task-path optimization, multi-trip harvesting schedules, robot-UAV collaboration, multi-orchard operation management, and the use of LLMs and VLMs in robotic task planning. Existing research suggests that single-robot systems are insufficient for large-scale orchard operations. Although multi-robot systems can alleviate efficiency bottlenecks, they also require effective solutions to task allocation, path planning, load balancing, and energy constraints. Accordingly, this article aims to provide a clear, comprehensive, and engineering-oriented review framework for intelligent orchard robots and to establish a basis for the subsequent analysis of environmental perception, autonomous navigation, precision operation, and collaborative decision-making. To illustrate the hierarchical structure of the technological system and the relationships among technical pathways and operation applications, Table 1 summarizes the core content, representative methods, supported tasks, and primary functions of each technological layer.

Table 1. Key technology framework and representative application tasks of intelligent orchard robots.

For the purpose of this review, a smart orchard robot is defined as an embodied or physically deployed robotic system designed to operate in orchard environments and capable of implementing at least part of an integrated perception–decision–action loop. Such systems may perform autonomous or semi-autonomous functions related to orchard inspection, navigation, harvesting, spraying, pollination, weeding, transportation, or coordinated task execution.

Ground mobile robots constitute the primary focus of this review. Aerial robots are included only when they directly perform orchard operations or cooperate with ground robots in sensing, task planning, scheduling, or execution. Fixed sensing platforms are not regarded as robots but are considered enabling infrastructure when their outputs are directly integrated into robotic perception, localization, planning, or control. Agricultural machinery is included only when it incorporates orchard-specific autonomous sensing, localization, decision-making, or operation-control capabilities.

Individual sensors, perception algorithms, manipulators, end effectors, and control modules are treated as robotic enabling subsystems rather than independent orchard robots. They are included when they have a demonstrated or clearly defined role within a robotic operation pipeline. Similarly, LLMs, VLMs, and other foundation models are considered only when they directly support robot perception, semantic reasoning, task planning, behavior generation, or control. General-purpose artificial-intelligence applications, conventional mechanized equipment, standalone remote-sensing studies, and broader smart-farming management systems without a direct connection to orchard-robot operation are outside the principal scope of this review.

To distinguish platform types from functional technologies, the reviewed studies are classified along two dimensions:

(1)
a platform dimension comprising ground robots, aerial robots, heterogeneous ground–aerial systems, and enabling robotic subsystems; and

(2)
a functional dimension comprising environmental perception, localization and navigation, precision operation execution, collaborative decision-making, and system-level evaluation.

To improve the transparency and reproducibility of this review, a structured literature search and study-selection process was conducted. The search was completed on 15 June 2026 using SPIS, covering publications from 2002 to 15 June 2026. Search terms covered orchard environments, robotic platforms, and key functions, including perception, navigation, mapping, harvesting, spraying, pollination, weeding, manipulation, multi-robot systems, robot–UAV cooperation, and vision-language models. Search syntax was adapted for each database, and reference lists of relevant reviews and primary studies were also examined.

A PRISMA-type procedure was used for study selection. The initial search identified 215 records. After duplicate removal, titles, abstracts, and full texts were screened. Studies were included if they investigated orchard robots or enabling technologies with a clear role in a robotic perception–decision–execution pipeline and reported sufficiently detailed or quantitative results. Studies focused solely on conventional machinery, standalone remote sensing, general smart-farming management, or algorithms without a clear robotic application were excluded. Editor