Performance Evaluation of Robotic Harvester with Integrated Real-Time Perception and Path Planning for Dwarf Hedge-Planted Apple Orchard

Abstract

Apple harvesting faces increasing challenges owing to rising labor costs and the limited seasonal workforce availability, highlighting the need for robotic harvesting solutions in precision agriculture. This study presents a 6-DOF robotic arm system designed for harvesting in dwarf hedge-planted orchards, featuring a lightweight perception module, a task-adaptive motion planner, and an adaptive soft gripper. A lightweight approach was introduced by integrating the Faster module within the C2f module of the You Only Look Once (YOLO) v8n architecture to optimize the real-time apple detection efficiency. For motion planning, a Dynamic Temperature Simplified Transition Adaptive Cost Bidirectional Transition-Based Rapidly Exploring Random Tree (DSA-BiTRRT) algorithm was developed, demonstrating significant improvements in the path planning performance. The adaptive soft gripper was evaluated for its detachment and load-bearing capacities. Field experiments revealed that the direct-pull method at 150 mN·m torque outperformed the rotation-pull method at both 100 mN·m and 150 mN·m. A custom control system integrating all components was validated in partially controlled orchards, where obstacle clearance and thinning were conducted to ensure operation safety. Tests conducted on 80 apples showed a 52.5% detachment success rate and a 47.5% overall harvesting success rate, with average detachment and full-cycle times of 7.7 s and 15.3 s per apple, respectively. These results highlight the system’s potential for advancing robotic fruit harvesting and contribute to the ongoing development of autonomous agricultural technologies.

1. Introduction

Apples are one of the most widely cultivated and economically significant fruits, with global production exceeding 90 million tons in 2022, primarily led by China, the United States, India, and Europe [1]. The production scale presents significant challenges in enhancing harvesting efficiency and orchard management. Traditional manual harvesting methods are labor-intensive, costly, and inefficient and rely heavily on the limited seasonal workforce availability, ultimately constraining orchard profitability and sustainability. Apple harvesting robots, a key advancement in agricultural automation, integrate advanced perception, intelligent planning, and efficient execution to enhance precision and efficiency in orchard management [2]. Recent studies have systematically explored system architectures, sensing technologies, fruit detachment mechanisms, and the remaining technical challenges and research opportunities in this domain [3,4,5].

Hedge-planting, an emerging method for cultivating dwarf apple trees in commercial orchards, features compact structures, optimized spacing, and a planar layout, which is ideal for high-degree-of-freedom robotic arms. By growing trees along trellises, it maximizes the orchard space, simplifies management, and streamlines harvesting operations [6]. Compared to other orchard architectures, such as V-trellis systems and spindle-shaped trees, it also improves the light distribution, stabilizes yields, and maintains low tree heights, creating a standardized, planar structure that reduces operational complexity [7]. The open canopy design reduces fruit occlusion, enabling sensors to detect apples rapidly and accurately, thereby improving the harvesting efficiency and precision [8]. Additionally, this structured layout improves robotic arm maneuverability, mitigating path planning challenges from overlapping branches and making hedge-planting ideal for robotic harvesting.

Accurate apple detection is hindered by variable lighting conditions, fruit occlusion, and natural variations in apple shapes and sizes [9]. Early detection methods, such as edge detection and Hough transforms, enabled an effective basic identification but are highly susceptible to background interference and lighting inconsistencies, severely limiting their practical applicability [10]. To address these limitations, Zhang et al. (2024) [11] proposed an improved You Only Look Once (YOLO)v5 model incorporating attention mechanisms and an optimized Intersection over Union (IoU) metric, resulting in a 10.9% improvement in the mean Average Precision (mAP). Similarly, Rathore et al. (2023) [12] combined YOLOv7 with EfficientNet-B0, achieving 90.2% detection precision and a 92.22% classification accuracy. Wang et al. (2025) [13] proposed an improved YOLOv8n model that enhanced the mAP by 2.9% while increasing the detection speed by 30.5%. Cao et al. (2025) [14] developed an improved lightweight YOLOv7-Tiny model tailored for complex orchard environments, reporting a 4% gain in the mAP. Additionally, generative adversarial networks have been utilized effectively to reconstruct partially occluded fruit images, thereby restoring textures and structural details to enhance detection accuracy [15]. However, scaling solutions to achieve lightweight, real-time processing remains a challenge [16].

Recent research has explored a range of path planning methods for robotic fruit harvesting, demonstrating a promising performance across different orchard environments. In spindle-shaped orchards, Li et al. (2023) [2] developed a Markov game-based task planning method for a three-joint, multi-arm harvester, achieving a 71.28% to 80.45% success rate with an average harvesting cycle of 5.8 s per apple. In complex unstructured orchards, Kang et al. (2024) [17] proposed an optimized Rapidly Exploring Random Tree (RRT) algorithm with adjustable expansion directions and dynamic step sizes for a 17-degree-of-freedom (DOF) humanoid robot, significantly improving the planning time, path length efficiency, and overall success rates. Similarly, Yan et al. (2024) [18] advanced the ant colony optimization algorithm by integrating bio-inspired strategies and B-spline smoothing, achieving a 96% planning success rate and a 50% reduction in planning times in MATLAB (Version 2024a, The Mathworks Inc., Natick, MA, USA) simulations. In structured orchard systems such as V-trellis arrangements, Silwal et al. (2017) [5] developed a priority-based optimization framework using the Traveling Salesman Problem and inverse kinematics, achieving an 84% harvesting success rate. Effective path planning is critical for generating smooth, efficient joint motions that enable precise fruit targeting while minimizing excessive movement or abrupt transitions. However, to date, no specific path planning approaches have yet been reported or validated for robotic harvesting in dwarf hedge-planted orchards.

End-effectors are crucial for harvesting efficiency, fruit integrity, and adaptability in complex orchards. They are categorized into gripping, suction-based, and flexible hybrid types, each offering unique advantages while addressing distinct challenges [19]. Gripping end-effectors use mechanical claws for fruit detachment. For instance, Au et al. (2023) [20] developed the Monash Apple Retrieving System, which combined a soft gripper with an optimized planning algorithm, achieving a 62.8% success rate and a median cycle time of 9.18 s per apple in unmodified orchard conditions. Silwal et al. (2017) [5] developed a tendon-driven, three-finger gripper employing “horizontal” and “tilted” pulling techniques, achieving 84.6% success. However, its effectiveness was limited with irregularly shaped or clustered apples. Suction-based end-effectors, which use vacuum or airflow mechanisms, have also been explored. Hu et al. (2022) [21] developed a suction-based device using a “rotational pulling” approach, achieving 78% success in controlled environments but only 47.37% in real-world orchards owing to environmental variability. Flexible hybrid end-effectors integrate gripping and suction mechanisms for greater adaptability. Wang et al. (2023) [22] proposed a hybrid model with flexible fingers and suction cups, incorporating active extension, compliance, and twisting modes, achieving 75.86% success while limiting fruit damage to 4.55%. However, in practice, suction cup slippage and tube clogging, particularly in humid or contaminated conditions, reduce reliability [23,24]. Despite notable advancements, high costs and integration complexity hinder the large-scale deployment of end-effectors in commercial orchards, thereby slowing the adoption of robotic harvesting systems. In light of these considerations, mechanical gripping end-effectors offer distinct advantages in terms of structural simplicity, cost-effectiveness, operational reliability, and compatibility with existing harvesting platforms. Therefore, this study adopts a gripping-type end-effector to enhance the harvesting efficiency and operational stability in complex orchard environments.

Overcoming these challenges requires advancements in real-time visual perception, efficient path planning algorithms, and cost-effective, multifunctional end-effectors to enable widespread robotic apple harvesting. Therefore, this study aims to develop a highly efficient and adaptable robotic apple harvesting system through the following key objectives: (1) deploying a lightweight YOLOv8n model to achieve an optimal balance between accuracy and speed for real-time apple detection in orchard environments; (2) enhancing the Bidirectional Transition-Based Rapidly Exploring Random Tree (BiTRRT) algorithm for an efficient and adaptive path planning framework; (3) refining detachment mechanisms and torque settings for adaptive soft grippers to ensure secure fruit detachment and load-bearing stability; and (4) integrating and validating the complete system through field testing to assess the combined performance of the visual perception, path planning, and gripper functionality in executing the sequential tasks of the fruit detection, approach, detachment, and placement in a dwarf hedge-planted apple orchard.

2. Materials and Methods

2.1. Orchard Environment

This study was conducted in an apple orchard managed by the Gangwon Agricultural Research and Extension Services in Gangwon-do, Republic of Korea. The orchard features a greenhouse equipped with an automated, intelligent adjustable lighting control system and adopts a modern hedge-planting pattern designed to enhance mechanized harvesting and overall productivity. Figure 1a shows the orchard layout containing trees approximately 2.5 m tall, spaced uniformly at 2–3 m between trees and rows. This layout maximizes spatial efficiency and facilitates the seamless operation of agricultural equipment. Figure 1b illustrates each hedge-planted tree with a horizontal trunk diameter of 3 cm, supporting 8 to 12 vertical branches approximately 1 cm in diameter. These branches are secured with taut support strings at 0.5 m, 1 m, and 1.5 m heights, ensuring structural stability, uniform fruit distribution, and consistent growing conditions. Apples grow within 0.5 m to 1.5 m in height, positioned on open branches with sparse foliage and clearly visible stems. This arrangement minimizes obstructions, optimizes fruit visibility, and streamlines the harvesting process. Support strings confine fruit growth to accessible zones, while wide inter-row pathways improve equipment mobility for efficient automated harvesting. The apples generally range from 6 to 9 cm in diameter, a size well-suited to the design specifications of the robotic gripper. Figure 1c illustrates a representative measurement.

2.2. Robotic System Design

2.2.1. Hardware System

The apple harvesting robotic system integrates advanced technologies for efficient, precise, and stable orchard operations. Figure 2 shows the system, mounted on the AgileX Bunker Unmanned Ground Vehicle (UGV) platform (AgileX Robotics, Shenzhen, China) and equipped with an Elite CS66 6-DOF robotic arm (Elite Robots, Suzhou, China) controlled by the ERB2E2K0-220/110 controller (Elite Robots, Suzhou, China). This robotic arm supports a 6 kg payload, provides a 914 mm working radius, and achieves high-precision positioning with ±0.02 mm repeatability. The end-effector features a Flexible Adaptive Servo Four-Finger Robotic Gripper (Easy Gripper, Dongguan, China) with a 19.4 cm length, designed to securely grasp apples while minimizing stem damage. The gripper operates through RS-485 communication, enabling seamless coordination with the robotic arm for optimized fruit handling. For visual perception, an Intel RealSense D435i RGB-D camera (Intel Corporation, Santa Clara, CA, USA) is mounted 10 cm above the gripper. This camera captures high-resolution (1280 × 720 pixels) RGB images and depth maps at up to 90 fps, with a wide field of view (87° × 58° × 95°). This facilitates precise 3D apple localization for accurate grasping and detachment. A laptop running Ubuntu 22.04 controls the system, interfacing with the robotic arm and camera through a TCP/IP network for real-time data processing and decision-making. A Thunder RS1000 emergency power unit (ROMOSS Technology, Shenzhen, China) with a 933 Wh ternary lithium (nickel–cobalt–manganese, NCM) battery provides continuous AC power to the robotic arm controller, which consumes approximately 150 W during operation and supports up to six hours of continuous use per charge.

2.2.2. Software System

The robotic apple harvesting system employs a modular task control process that seamlessly integrates perception, planning, and execution for precise localization, efficient path planning, and reliable harvesting in dynamic orchards. Built on the Robot Operating System (ROS) 2 humble framework, the system leverages real-time communication, multithreading, and a modular architecture to ensure stability under complex field conditions. Figure 3 illustrates the system workflow comprising the following three main stages: (1) Perception: At the beginning of each harvesting cycle, the robotic arm remains in its initial position, while the RGB-D camera captures high-resolution RGB and depth data. Building on our previously established detection and localization framework by Jin et al. (2025) [25], this study incorporates a newly developed recognition model to identify apples and estimate their 3D coordinates. A separate quantitative assessment of localization accuracy was not conducted in this study. An eye-in-hand transformation then converted these coordinates into the base frame of the robotic arm, following methodologies established by Song et al. (2023) [26] and Yan et al. (2023) [27]. (2) Path Planning: The system selects an optimal target point based on reachability and feasibility, ensuring safe motion within joint limits, workspace constraints, and obstacles. MoveIt!2 utilizes the Open Motion Planning Library (OMPL) to generate motion trajectory, while inverse kinematics are computed using the Kinematics and Dynamics Library (KDL), a Jacobian-based numerical solver. Joint constraints are managed internally during the planning process, allowing for stable and efficient real-time motion adjustments in 6-DOF robotic manipulators. (3) Harvest Control: Prior to grasping, the system verifies camera and robotic arm functionality without capturing additional images. Upon detecting a reachable apple, the end-effector aligns, opens the gripper, and follows the planned trajectory to the pre-grasp position. The adaptive soft gripper then conforms to the apple’s shape and applies controlled pressure—regulated by servo motor torque commands from the host computer—to detach the fruit via a twisting or pulling motion. After successful detachment, the robotic arm moves to the pre-placement position, releases the apple at the drop-off point, and resets for the next harvesting cycle.

2.3. Apple Perception

2.3.1. Dataset and Model Training

This study utilizes publicly available datasets from Santosh (2019) [28] and Liu et al. (2024) [29], containing captured orchard images under diverse conditions, including variations in apple maturity, lighting, viewing angles, and occlusion. All images were annotated, standardized in TXT format, and manually verified using the labeling tool [30]. To align with the camera perspective of the robot, only clear and relevant images with accurately labeled apples were retained. The final dataset comprises 4709 images, with approximately 46,173 labeled apples. Following established data partitioning methods [25,31,32], the dataset was randomly split into 3297 training, 942 validation, and 470 testing images in a 7:2:1 ratio. Extensive experiments and hyperparameter tuning were conducted to optimize detection performance. All models were trained under identical conditions for fair comparison. Input images were resized to 640 × 640 pixels, and training spanned 300 epochs. Optimization was performed using the Stochastic Gradient Descent algorithm, with a 0.01 learning rate, 0.937 momentum coefficient, and a batch size of 32 images. Model training was performed on a high-performance workstation running Windows 11 Professional (version 23H2, build 22631.4460), equipped with an Intel^® Core™ i9-14900KF CPU @ 3.20 GHz, 96 GB of RAM, and an NVIDIA RTX A5000 GPU (24,564 MiB VRAM) for computational acceleration. The models were implemented in PyTorch2.2.1 and trained under consistent experimental conditions to ensure reproducibility and reliability.

The performance of object detection models was evaluated using precision and recall metrics. Precision measures detection accuracy by quantifying the proportion of correctly predicted positives among all predicted positives. Conversely, recall assesses the proportion of correctly identified positives among all actual positives, reflecting the detection capability of the model. Additional factors such as parameter count (M) for model size and complexity, floating-point operations (FLOPs, G) for computational cost, and detection speed (milliseconds per image, ms/image) assess computational efficiency and real-time applicability. These metrics collectively determine the suitability of the model for real-world automated apple harvesting.

2.3.2. YOLO Improvement Strategy

YOLO, with its advanced feature representation and end-to-end learning capabilities, excels in detecting multiple and small objects in complex environments, making it ideal for real-time, precise apple harvesting in orchards [33]. Among various YOLO iterations, YOLOv8n is a reliable and widely adopted model [34,35,36,37], offering an optimal speed balance, lightweight architecture, simplicity, and performance in resource-constrained environments [38]. At the time of this study, YOLOv8 was the latest stable release, supported by mature implementations and an active development community. By optimizing its backbone, neck, and head architecture with the enhanced C2f module, YOLOv8n significantly improves detection performance [39]. To further improve real-time performance and detection accuracy, this study integrates the Faster module into the C2f structure of YOLOv8n, forming the lightweight C2f–Faster model. The C2f–Faster model maintains a compact model size while reducing computational overhead and has demonstrated promising results in previous detection tasks [40]. In this study, it is applied and evaluated for apple detection for the first time. Figure 4a depicts the C2f–Faster architecture consisting an initial convolution layer (Conv) for feature extraction, feature decomposition (Split) for diversified feature representation, multiple Faster_Block layers for enhanced feature refinement, feature fusion (Concat) to integrate multi-scale information, and a final output convolution layer for detection [41]. The Faster_Block optimizes feature processing through the following: Channel adjustment (Conv1x1): Reduces and expands dimensions efficiently; Local feature modeling (Partial Conv3): Extracts spatial features with high efficiency; Channel interaction (MLP): Enhances feature representation across channels; Regularization (DropPath): Dynamically regulates feature paths, improving model robustness and generalization; and Residual connections (Shortcut): Maintains gradient flow stability and input integrity. Unlike the original Bottleneck module, the C2f–Faster model employs Partial Conv3 for efficient spatial modeling and utilizes dual Conv1x1 layers for optimized channel compression and expansion. These enhancements reduce computational overhead while enhancing inter-channel interaction [42]. Additionally, DropPath stabilizes features, while residual connections maintain gradient consistency, enabling robust performance in complex orchards.

Figure 4b illustrates the C2f–Faster module integrated into the backbone and neck, forming the YOLOv8n–C2f–Faster model used in this study. This integration enhances feature extraction through advanced decomposition and the inclusion of stacked Faster module layers. The streamlined architecture is designed to optimize feature representation for robust and reliable performance in complex and dynamic orchard conditions. The YOLOv8n–C2f–Faster model ensures real-time rapid and accurate apple detection while simultaneously generating precise grasping coordinates crucial for orchard harvesting. The lightweight design of the C2f–Faster module significantly reduces computational overhead, enhancing detection speed and robotic arm efficiency in field conditions. Furthermore, the compact architecture of the model and reduced resource demands enable the deployment of edge devices and embedded systems, facilitating seamless integration into autonomous agricultural machinery.

2.4. Apple Picking Path Planning

2.4.1. BiTRRT Algorithm

While RRT offers simplicity and rapid exploration, it often generates suboptimal and inefficient paths owing to its lack of built-in optimization [43]. Rapidly Exploring Random Tree Connect (RRTConnect) enhances planning speed through bidirectional tree expansion to rapidly connect start and goal nodes. However, it remains prone to path redundancy, increased energy consumption, and limited adaptability in dynamic environments. BiTRRT addresses these limitations by incorporating dynamic temperature control and transition tests, significantly improving path quality and efficiency. The temperature control mechanism adjusts exploration strategies based on the cost of newly generated paths, where the “temperature” denotes a probabilistic parameter that allows the acceptance of higher-cost solutions to escape local minima during optimization [44]. Transition tests further refine this process by guiding the search toward cost-effective regions, mitigating inefficiencies in traditional RRT methods. BiTRRT provides a well-balanced trade-off between path optimality and computational load, resulting in smoother, more efficient trajectories. Its intelligent balance between exploration and exploitation reduces redundancy and enhances adaptability, making it particularly effective in complex and dynamic environments [45,46]. Algorithm 1 presents the detailed pseudocode for the BiTRRT algorithm.

Algorithm 1: Pseudocode of the BiTRRT algorithm

Input: Configuration space C, cost function c: C → ℝ+, initial configurations qinit, goal configurations qgoal. Output: Path from qinit to qgoal, or failure if no valid path exists. 1:  T1 ← Init tree (qinit), T2 ←Init tree(qgoal) 2:  Temp ← Init temperature 3:  Cost threshold ← Constant value 4:  While not converged: 5:   qrand ← Sample random configuration(C) 6:   $q_{near}^1$ ← Find the nearest node (T1, qrand) 7:   If refinement control (T1, $q_{near}^1$, and qrand): 8:    qnew ← Extend ($q_{near}^1$, qrand) 9:    If qnew ≠ null: 10:     If $\mathit{c}\left(q_{new}\right)$ ≤ cost threshold: 11:      Add node and edge (T1, $q_{near}^1$, and qnew) 12:      $q_{near}^2$ ← Find the nearest node (T2, qnew) 13:       Attempt link (T1, qnew, T2, and $q_{near}^2$) 14:   Swap (T1, T2) 15: Return success if trees connect; otherwise, failure

The BiTRRT algorithm employs a bidirectional search framework, simultaneously constructing two search trees, T₁ and T₂, from the start point $q_{init}$ and goal point $q_{goal}$, respectively. Using random sampling at $q_{rand}$ and iterative path extension, the algorithm efficiently explores the configuration space. As its core, BiTRRT uses a path cost function c: C→$\mathbb{R}+$ to evaluate traversal configuration cost within space C. This cost-based optimization dynamically adjusts the search process to prioritize cost-efficient regions while minimizing unnecessary detours. At each extension step, the algorithm performs the following sequence: (1) a random sample $q_{rand}$ is generated within the configuration space, and (2) the nearest node $q_{near}^1$ in the tree is identified using a nearest-neighbor search, and the path then extends from $q_{near}^1$ to $q_{new}$, forming a potential new node $q_{new}$. The acceptance of $q_{new}$ is determined by the transition probability P, which depends on the cost difference between nodes and the current temperature of the system. This acceptance criterion, formally defined in Equations (1) and (2), prioritizes lower-cost paths while allowing exploration of suboptimal routes when necessary to avoid local minima.

$$∆c=c \left(q_{new}\right)-c \left(q_{near}^1\right)$$

(1)

$$P=\exp \left(-{\displaystyle \frac{∆c}{Temp}}\right)$$

(2)

where $c\left(q_{new}\right)$ is the cost of the newly extended node, while $c\left(q_{near}^1\right)$ denotes the cost of the nearest node in the current tree, and $∆c$ represents the change in path cost per extension step, serving as a critical metric for evaluating the efficiency of potential paths. Temp represents the temperature, which is initially set to enable broad exploration and gradually decreases through exponential or linear decay as the algorithm progresses, regulating path acceptance.

Additionally, the algorithm incorporates a fixed cost threshold, $c\left(q_{new}\right)$< cost threshold, to further restrict the exploration of excessively costly paths. This ensures that the search remains focused on feasible and cost-efficient trajectories, avoiding unnecessary computational effort on suboptimal routes. Once $q_{new}$ successfully meets both the transition test and cost threshold, it integrates into the current search tree. Subsequently, the algorithm identifies the nearest node $q_{near}^2$ in the opposing tree T₂ and attempts to connect the two trees. If a successful connection is established, BiTRRT returns a low-cost path from the start to the goal, optimized for both efficiency and robustness. If it fails, the algorithm resumes random sampling and path extension, until meeting a predefined termination condition—reaching a maximum number of iterations, time constraints, or successfully finding an optimal path.

2.4.2. BiTRRT Improvement Strategy

Although BiTRRT is effective for general path planning, it faces specific limitations when applied to real-time dynamic robotic apple picking. In this study, the orchard environment was optimized by removing physical obstacles, reducing spatial constraints, and enabling the algorithm to prioritize efficient path planning over extensive collision avoidance. While these modifications have successfully eliminated collision-related redundancies, several challenges persist, including fixed temperature adjustments, overly rigid or simplistic cost threshold evaluations, and computationally intensive transition probability models. These issues contribute to redundant exploration, inefficiencies, and limited adaptability, particularly in environments requiring rapid and precise decision-making. To overcome these limitations, this study employs an enhanced algorithm introduced as Dynamic Temperature Simplified Transition Adaptive Cost Bidirectional Transition-Based Rapidly Exploring Random Tree (DSA-BiTRRT). Algorithm 2 presents the pseudocode of the enhanced algorithm.

Algorithm 2: Pseudocode of the DSA-BiTRRT algorithm

Input: Configuration space C, cost function c: C → ℝ+, initial configurations qinit, goal configurations qgoal. Output: Path from qinit to qgoal, or failure if no valid path exists. 1:  T1 ← Init tree (qinit), T2 ←Init tree(qgoal) 2:  Temp ← Init temperature, 3:  Best cost ← ∞, worst cost ← 0 4:  While not converged: 5:   qrand ← Sample random configuration(C) 6:   $q_{near}^1$ ← Find the nearest node (T1, qrand) 7:   If refinement control (T1, $q_{near}^1$, and qrand): 8:    qnew ← Extend ($q_{near}^1$, qrand) 9:    If qnew ≠ null: 10:     Update the best cost and worst cost based on c (qnew) 11:     Cost threshold ← best cost + α ·(worst cost − best cost) 12:     If c (qnew) ≤ Cost threshold: 13:       Add node and edge (T1, $q_{near}^1$, and qnew) 14:       $q_{near}^2$ ← Find the nearest node (T2, qnew) 15:      Attempt link (T1, qnew, T2, and $q_{near}^2$) 16:   Swap (T1, T2) 17: Return success if trees connect; otherwise, failure

The DSA-BiTRRT algorithm incorporates three key enhancements—dynamic temperature adjustment, a simplified transition probability model, and an adaptive cost threshold. The details are as follows:

• (1) Dynamic temperature adjustment: Fixed temperature settings in the original BiTRRT fail to adapt to fluctuating path costs in dynamic environments, resulting in inefficient exploration and suboptimal paths. The temperature parameter Temp is continuously adjusted based on the best and worst observed path costs, as defined in Equation (3).

  $$Temp=\max \left(Init temperature·\left(1-{\displaystyle \frac{best cost}{worst cost}}\right),\epsilon \right)$$

  (3)

  where Init temperature is the initial temperature setting, best cost and worst cost represent the lowest and highest path costs encountered, and $\epsilon$ is a small constant that prevents Temp from reaching zero. This approach progressively lowers the acceptance probability of high-cost paths as better paths emerge, thereby improving efficiency, minimizing redundant exploration, and enhancing the adaptability of the algorithm to dynamic orchard environments.
• (2) Simplified linear transition probability: The original BiTRRT employs an exponential model for balancing exploration and exploitation; however, its high computational costs, particularly in high-dimensional spaces, limit real-time efficiency. To address this, a simplified linear model replaces the exponential approach, significantly reducing computational complexity while preserving path optimization. The transition probability P is now defined by the simplified linear equation in Equation (4), enhancing real-time performance.

  $$P=1-{\displaystyle \frac{c\left(q_{new}\right)}{Temp}}$$

  (4)

  where $c\left(q_{new}\right)$ represents the cost of the newly extended mode, and Temp is the dynamically adjusted temperature from Equation (3). This modification enhances the adaptability of the model to dynamic environments while prioritizing cost-efficient paths, thereby optimizing real-time agricultural automation.
• (3) Adaptive cost threshold: The original BiTRRT relies on a fixed cost threshold to filter high-cost paths, which can either overly restrict exploration or allow inefficient paths, limiting efficiency and adaptability in dynamic environments. To address this, the cost threshold is dynamically updated based on the best and worst observed path costs. This adaptive approach refines exploration, directing the algorithm toward promising regions while reducing computational overhead. The adaptive cost threshold is defined in Equation (5).

  $$Cost threshold=best cost+\alpha ·\left(worst cost-best cost\right)$$

  (5)

  where $\alpha$ is a sensitivity parameter that modulates the rate of threshold adjustments. This adaptive mechanism allows the algorithm to dynamically adjust to changing conditions by prioritizing optimal paths, reducing redundant computations, and enhancing path quality, thereby ensuring efficient and flexible planning for real-time agricultural automation.

2.5. Apple Picking Control

The apple-picking robot, built on the ROS2 framework, seamlessly integrates detection, localization, motion control, and coordination to achieve precise and efficient apple harvesting. This modular approach synchronizes perception and actuation, optimizing the performance of the robot in dynamic orchard environments. At the beginning of each picking cycle, the robotic arm moves to its predefined initial position, activating the target detection model and automatically opening the gripper to signal the commencement of a new task. The system captures real-time 3D spatial coordinates of apples within the workspace using an onboard camera. Rather than adopting a complex multi-target strategy, the robot focuses on picking the last detected apple within its reachable area each cycle, simplifying control logic while maintaining efficiency. Once the position of the apple is detected and published, the robotic arm controller subscribes to the position message and generates an optimal trajectory through motion planning. The arm initially moves to a pre-grasp position for proper alignment, then performs a precise linear approach toward the apple. Upon reaching the target, the system instructs the adaptive gripper to close, securely completing the picking operation. After successfully grasping the apple, the arm retracts and moves to a pre-designated placement position above the designated drop zone. During this motion, posture adjustments are automatically managed through the motion planning process to ensure apple stability and prevent slippage. The gripper then releases the apple, ensuring precise placement at the designated location. To maintain robust system performance, critical timing data is recorded at key stages, including the detection-to-motion initiation delay, grasping completion time, and total picking cycle duration. After completing each picking cycle, the arm resets to its initial position, reactivating the detection model to initiate the next cycle. A built-in timing mechanism monitors delays, while a retry function addresses motion planning failures, ensuring continuous, robust, and efficient apple picking operations in dynamic orchard environments. Algorithm 3 shows the pseudocode for the apple picking control process.

Algorithm 3: Pseudocode for the apple picking control

1: Initialize system: 2:   Movej (home position) → Reset robotic arm to initial position. 3:   Activate the detection model. 4:   Gripper state = open. 5: While true do: 6:   Detect target: 7:    Capture apple positions with a camera. 8:    Identify the last apple in the workspace. 9:    Publish the 3D coordinates of the apple to ROS2. 10:   Motion planning and execution: 11:    Subscribe to the target position. 12:    Movej (pre-grasp position) → Move plan trajectory near the target. 13:    Movel (grasp position) → Linearly approach the target. 14:   If work state = true (grasping phase): 15:    Gripper state = closed. 16:   Else (placing phase): 17:    Movel (placement position) → Move plan trajectory to placement. 18:    Adjust posture during motion. 19:    Gripper state = open. 20:   Logging and preparation: 21:    Record timing data for detection, grasp, and cycle duration. 22:    Movej (home position) → Return robotic arm to initial position. 23:    Reactivate the detection model. 24: End loop if no valid target.

2.6. Experiment Setup

Figure 5a shows that the robotic arm was securely mounted on a laboratory tabletop to harvest artificial apples, each with a diameter of 8 cm, suspended from a simulated vine. This study commenced with the validation and optimization of target perception and path planning in a controlled laboratory setting. Following successful laboratory trials, a week-long field experiment was conducted in early November 2024 under clear weather conditions, coinciding with the peak harvest maturity period. In the field experiments shown in Figure 5b, apple branches were fixed to wires (Figure 1b). To simplify laboratory and field experiments, complex grasping strategies and obstacle detection algorithms were excluded. During field trials, additional safety measures were incorporated to prevent unintended collisions between the robotic arm or gripper and obstacles such as branches, support wires, or other non-target objects. The UGV was manually repositioned between trees or row segments to locations where apples were visible and within the manipulator’s reach. There was no predetermined positioning interval; instead, placement was determined by fruit distribution and in-field conditions. Obstacles within a 5 cm radius of the target apples were removed, and thinning operations were conducted to create sufficient clearance for safe robotic operation. As a result, repeated repositioning and overlapping observation areas were generally unnecessary. In the field trials, the harvesting force and load-bearing capacity of the adaptive soft gripper necessary for successful apple detachment were also assessed, comparing various detachment methods (e.g., twisting vs. pulling). Finally, the entire robotic system underwent integrated testing to validate its operational functionality and harvesting capability under modified field conditions.

3. Results and Discussion

3.1. Performance of Different Perception Algorithms

A comprehensive evaluation was conducted to compare the proposed YOLOv8n–C2f–Faster network with established models, including YOLOv3-tiny, YOLOv6, and the original YOLOv8n.

Table 1. Performance comparison of object detection algorithms, evaluating precision, recall, parameter count, FLOPs, and detection speed to assess efficiency and accuracy.

Model	Precision (%)	Recall (%)	Parameters (M)	FLOPs (G)	Speed (ms/Image)
YOLOv3-tiny	88.45	85.23	12.13	18.90	9.70
Yolov6n	92.78	75.85	4.23	11.80	9.60
YOLOv8n	93.06	74.72	3.01	8.10	9.20
YOLOv8n–C2f–Faster	93.66	70.98	2.30	6.30	8.90

shows a summary of the recognition performance of these networks on the test dataset1, detailing the precision, recall, parameter count, FLOPs, and detection speed. The YOLOv8n–C2f–Faster model achieved the highest precision at 93.66%, significantly surpassing the other models and indicating a significantly low false positive rate. This precision exceeded that of YOLOv3-tiny, YOLOv6, and YOLOv8n by 5.21%, 0.88%, and 0.60%, respectively, highlighting its superior ability to minimize incorrect detections. Regarding model complexity and computational efficiency, YOLOv8n–C2f–Faster demonstrated a clear advantage with its lightweight architecture. It featured the lowest parameter counts (2.30 M) and FLOPs (6.30 G) among all compared models, highlighting its computational efficiency. Its parameter count was 23.9% and was 81% lower than that of YOLOv8n and YOLOv3-tiny, respectively, significantly reducing memory requirements and accelerating the inference speed. Furthermore, YOLOv8n–C2f–Faster delivered the fastest detection speed, processing each image in an average of 8.9 ms. This speed was 0.30, 0.70, and 0.80 ms faster than that of YOLOv8n, YOLOv6, and YOLOv3-tiny, respectively, demonstrating its capability for real-time performance in resource-constrained environments. Although its recall was relatively lower at 70.98%, suggesting a potential for missed detections owing to factors such as high confidence thresholds, complex backgrounds, and limited training data diversity, the overall performance of the model remained highly competitive. Its exceptional efficiency is attributed to the enhanced C2f–Faster architecture, which incorporates optimized modifications to the YOLOv8n backbone and neck. These architectural improvements effectively reduced the computational overhead while maintaining a high detection accuracy, meeting the stringent real-time and efficient requirements of orchard harvesting operations.

In conclusion, this study demonstrates the effectiveness of YOLOv8n–C2f–Faster for apple detection and its promising potential for integration into robotic arms used for orchard harvesting. The model achieves an optimal balance between precision and efficiency, with its lightweight design and minimal computational demands enabling a seamless deployment on robotic platforms. This makes it well-suited for real-time agricultural applications, where accuracy and speed are essential.

3.2. Performance of Different Path Planning Algorithms in the Laboratory

A simulated harvesting experiment was conducted in a controlled laboratory environment, utilizing a fixed artificial apple target to replicate the path planning and execution processes involved in real-world harvesting. Figure 6 illustrates a successful apple harvesting sequence in the laboratory, detailing each step from the initial detection and localization to the final placement. The process began with the camera-based detection, localization, and gripper opening, followed by the robotic arm executing the planned detachment path from the initial position to the re-grasping point before proceeding directly to the pre-grasp position. The gripper securely grasped the apple, applied a direct pulling motion for detachment, and then moved to the pre-placement position. Finally, the gripper released the apple into the designated storage area. In cases where path planning failed—due to inaccessible target positions or transient environmental changes—the robotic arm returned to its initial state and initiated a new cycle of perception and planning. This built-in retry mechanism enables the system to maintain task continuity and robustness, even in the presence of occasional planning failures.

The proposed DSA-BiTRRT algorithm was evaluated based on efficiency, path quality, and robustness during the detachment and placement stages of the harvesting process. To comprehensively assess its effectiveness, a comparative analysis was conducted against classical path planning algorithms, including RRT, RRTConnect, Transition-Based Rapidly Exploring Random Tree (TRRT), and BiTRRT. The path planning efficiency was measured based on the time required to generate the first executable path, excluding the execution time of the robotic arm. The results are shown in Figure 7. Each algorithm was executed 100 times under identical obstacle-free conditions to simulate grasping a fixed-position target in the laboratory, with a maximum planning time of 5 s per trial. Trails that failed to generate a feasible path within this timeframe were recorded as planning failures.

The experimental results show that nearly all algorithms successfully generated feasible paths within the specified time frame. As shown in

Table 2. Comparative performance evaluation of various path planning algorithms across 100 laboratory harvesting trials, assessing planning time for detachment and placement paths.

Path Planning Algorithm	Detaching Path Time (s)	Placing Path Time (s)	Average Planning Time (s)	Standard Error (s)
RRT	0.329	0.470	0.399	0.024
RRTConnect	0.297	0.360	0.328	0.007
TRRT	0.373	0.608	0.490	0.012
BiTRRT	0.194	0.294	0.244	0.003
DSA-BiTRRT	0.189	0.284	0.237	0.003

and

Table 3. Comparative performance evaluation of various path planning algorithms across 100 laboratory harvesting trials, assessing planning length and failure rates for detachment and placement paths.

Path Planning Algorithm	Detaching Path Length (m)	Placing Path Length (m)	Average Planning Length (m)	Standard Error (m)
RRT	0.744	1.627	1.186	0.003
RRTConnect	0.840	1.600	1.220	0.004
TRRT	0.757	1.602	1.180	0.002
BiTRRT	0.796	1.604	1.200	0.003
DSA-BiTRRT	0.780	1.589	1.184	0.002

, both the planning time and path length exhibited consistently low standard errors (ranging from 0.003–0.024 s to 0.002–0.004 m, respectively), reflecting a stable and reliable performance across repeated trials. The proposed DSA-BiTRRT achieved an average planning time of 0.189 s for the detaching path and 0.284 s for the placement path, yielding an overall mean planning time of 0.237 s. This resulted in a 2.87% reduction compared to that of the original BiTRRT, demonstrating a superior efficiency across all harvesting stages compared to that of RRT, RRTConnect, and TRRT, highlighting its effectiveness in optimizing path planning. For the path quality evaluation, the Euclidean distances of the joint coordinates of the robotic arm were summed to assess the optimization performance. The results showed that the proposed DSA-BiTRRT exhibited a stable and superior optimization, achieving an average path length that was 1.33% shorter than that of the original BiTRRT. Furthermore, it outperformed RRT and RRTConnect in path quality while maintaining an optimization level comparable to that of TRRT. In terms of robustness, the algorithm exhibited a failure rate of only 1% across 100 trials, causing an 80% reduction compared to the original BiTRRT. These results underscore the proposed DSA-BiTRRT’s high reliability, robust path planning capability, and consistent superiority over conventional algorithms.

The superior performance of DSA-BiTRRT is attributed to the integration of the dynamic cost threshold adjustment, dynamic temperature control, and a simplified linear transition probability mechanism. These advancements significantly enhance stability and efficiency, making the algorithm well-suited for robotic apple picking applications. By effectively optimizing the path quality, planning efficiency, and success rates, DSA-BiTRRT provides a robust and reliable solution for robotic arm path planning in apple harvesting. The findings validate its strong performance and broad applicability, establishing it as a significant advancement in agricultural robotics.

3.3. The Performance of the Adaptive Soft Gripper in the Field

The detachment force and the load-bearing capacity of the adaptive soft gripper in handling apples were key factors influencing the energy efficiency and operational performance. Before the formal experimentation, the critical torque threshold for the apple detachment was determined by incrementally adjusting the torque of the gripper using the rotation- and direct-pull methods. Two detachment approaches were evaluated under varying torque conditions: the rotation-pull method at 100 and 150 mN·m (with 180° rotation at 4 rad/s) and the direct-pull method at 150 mN·m.

Table 4. The performance comparison of the detachment and placement between the rotation-pull method (tested at 100 and 150 mN m) and the direct-pull method (tested at 150 mN m), evaluating their effectiveness in apple harvesting.

The Number of Apples	Torque (mN·m)	Detachment Method	Detachment Success Rate (%)	Placement Success Rate (%)
12	100	Rotation-pull	91.67%	0%
12	150	Rotation-pull	100%	100%
12	150	Direct-pull	100%	100%

summarizes the outcomes from trials conducted on 12 apple targets under optimal gripping conditions, with Figure 8 providing a visual representation using a 3D bar chart. The rotation-pull method at 100 mN·m achieved a 91.67% detachment success rate, with one failure attributed to branch elasticity. However, at this torque level, inertial forces during the detachment caused all apples to be flung away from the gripper, resulting in a placement success rate of 0%. Increasing the torque to 150 mN·m significantly improved the performance, achieving a 100% detachment rate while ensuring successful placement, leading to a 100% placement success rate. Although the rotation-pull method facilitated effective detachment, it introduced unnecessary operational complexity, reducing efficiency and increasing the energy consumption, while the direct-pull method streamlined the process by pulling apples directly along the placement path. At a torque of 150 mN·m, this method achieved the same 100% detachment and placement success rates as the rotation-pull method while reducing complexity, making it a more efficient and practical approach for robotic apple harvesting.

The experimental findings confirm the effectiveness and stability of the direct-pull method at a torque of 150 mN·m, highlighting its strong potential for improving detachment efficiency. Among the two evaluated methods, the direct-pull approach demonstrated not only consistent performances and reliability but also facilitated a simplified, linear motion of the robotic arm without rotational movement, making it a highly practical solution for orchard harvesting. However, while higher torque levels significantly increased detachment and placement success rates, they also increased the energy consumption and heightened the risks of damaging the apples and equipment. Achieving an optimal balance between the torque and load capacity was critical, as insufficient torque compromised the ability of the gripper to securely hold and place detached apples, whereas excessive torque, while improving grip strength, risked compromising the structural integrity of the system and the quality of the harvested produce. These findings indicate the importance of carefully optimizing torque settings to achieve an optimal balance between detachment effectiveness, energy efficiency, and operational safety.

3.4. The Integration and Field Harvesting Evaluation of the Autonomous Robotic Arm

A comprehensive apple harvesting experiment was conducted to assess the field performance of the robotic arm system, integrating the YOLOv8n–C2f–Faster algorithm for precise apple detection and localization, the DSA-BiTRRT algorithm for efficient path planning, and an adaptive soft gripper employing a pulling detachment strategy at 150 mN·m torque. The experiment targeted 80 apples positioned within the camera’s field of view in a controlled environment where obstacles were cleared to reflect typical variations in fruit positioning while minimizing the interference from the occlusion or clustering. Under these conditions, the system achieved an average detection and localization time of approximately 0.03 s. To prevent target loss due to a reduced camera field of view during close-range maneuvers, the fruit detection was performed solely at the initial position, without the camera input during the approach. Figure 9 illustrates detection results captured by the D435i camera during orchard trails, displaying bounding boxes, labels, confidence scores, and center points of the apples within the camera coordinate system. Figure 9a shows a single detected apple, while Figure 9b demonstrates the capability of the system to accurately identify and annotate multiple apples, even in complex environments. This approach enhanced real-time monitoring and improved decision-making during harvesting operations. If no graspable apples were detected within the current field of view, the UGV was manually repositioned to a new picking location to continue the harvesting process.

Figure 10 illustrates a successful apple harvesting sequence captured during the field trial, showing the controlled release of the apple into the storage box positioned at the rear of the UGV. To assess the system performance, key metrics—including the detection, grasping, detachment, placement, and overall harvesting [47]—were statistically analyzed. A summary of these results is presented in

Table 5. Summary of field harvesting experiment results, showing performance metrics for detection, grasping, detachment, placement, and overall harvesting success.

Process	Indicator	Value
Detection	Number of apples detected	80
Grasping	Number of apples grasped	45
	Grasping success rate (%)	56.3%
Detachment	Number of apples detached	42
	Detachment success rate (%)	52.5%
	Detachment average time (s)	7.7
Placement	Number of apples placed	38
	Placement success rate (%)	47.5%
	Placement average time (s)	14.8
Overall harvesting	Number of apples harvested	38
	Harvesting success rate (%)	47.5%
	Harvesting average time (s)	15.3

. All target apples were successfully detected, and 45 apples were securely grasped, resulting in a grasping success rate of 56.3%. During detachment, 42 apples were successfully detached, corresponding to a detachment success rate of 52.5% with an average detachment time of 7.7 s. During placement, 38 apples were accurately placed into the storage box, yielding a placement success rate of 47.5% and an average placement time of 14.8 s. The overall harvesting success rate was 47.5%, with an average cycle time of 15.3 s per apple. Upon the completion of each cycle, the system automatically reinitiated the detection and localization process, enabling the robotic arm to seamlessly transition to the next apple harvesting operation.

3.5. Failure Case Analysis

The results show that a significant number of apples were not successfully harvested, primarily attributed to challenging natural lighting conditions, the interference from branches and leaves, and limitations in the adaptability of the system. As shown in Table 5 and Figure 11, among the 80 apples targeted for harvesting, 42 (52.5%) were not harvested owing to failures occurring in the grasping, detachment, or placement stages, as detailed below.

Failures in the grasping stage accounted for the majority of unsuccessful harvest attempts, with 35 apples (83.3% of all failures) affected by one of four primary factors. A key limitation was the recall performance of the YOLOv8n–C2f–Faster model (70.98%), which when coupled with insufficient dataset diversity, led to perception failures under varying canopy structures and lighting conditions. Positioning errors were responsible for 21.4% of failures (nine apples), often occurring when target apples were partially obscured by foliage or affected by inconsistent illumination. These conditions hindered accurate localization, revealing the model’s sensitivity to occlusion and lighting variability. Unintended branch grasping was another major cause, accounting for 28.6% of failures (twelve apples). This issue was particularly common in hedge-planted orchards, where the close proximity of branches and short fruit stems increased the likelihood of the gripper inadvertently grasping twigs or colliding with trellis wires. These failures were exacerbated by the limited detection accuracy and the absence of adaptive gripper correction mechanisms. Path obstructions accounted for 31.0% of failures (thirteen apples), which stemmed from inadequate obstacle avoidance capabilities and the incomplete pre-harvest clearing of the system, leading to branch interference with the arm movement of the robot. Additionally, accidental apple dislodging (2.4% of failures, one apple) occurred when the robotic gripper inadvertently struck the apple, causing it to be dislodged before a secure grasp could be established. These failure cases collectively highlight limitations in the perception, motion planning, and end-effector control—specifically, the need for enhanced obstacle avoidance, improved environmental perception, and more adaptive manipulation strategies.

Failures during the detachment stage were relatively rare, accounting for only three apples (7.1% of all failures). Among these, two apples (4.8%) remained attached owing to branch elasticity, which absorbed the applied pulling force and prevented successful detachment. These cases primarily reflect the limitations of using a single detachment strategy under variable physical conditions. Additionally, one apple (2.4%) failed to detach owing to system malfunctions caused by hardware failures or algorithmic response delays, highlighting the need for improved hardware reliability and real-time fault-tolerant control.

Failures in the placement stage affected four apples (9.5% of all failures). Posture deviations (4.8%, two apples) led to unstable grips, causing apples to drop during the placement process. These failures were primarily attributed to perception inaccuracies that compromised the grasp stability. Similarly, path obstructions (4.8%, two apples) hindered successful placements when branches obstructed the movement of the robotic arm, highlighting continued limitations in motion planning and obstacle avoidance. Despite these issues, the placement return path was more direct and controlled, and the designated placement areas had fewer obstacles, leading to a lower failure rate compared to that of the grasping stage.

Overall, 35 of the 42 failed attempts (83.3%) resulted from positioning errors and inadvertent branch grasping. The further analysis revealed that these issues primarily stemmed from inherent camera inaccuracies, localization disruptions caused by orchard lighting variations, and the proximity of apples to branches, which led to the unintended simultaneous grasping of the apple and the surrounding vegetation. Additionally, the system lacked effective obstacle avoidance mechanisms, and the insufficient pre-harvest clearing allowed branches to interfere with the robotic arm during harvesting. These findings indicate the urgent need to enhance localization accuracy, improve adaptability to dynamic lighting conditions, and integrate advanced obstacle avoidance algorithms to significantly improve harvesting success rates.

3.6. Discussion

This study presents several novel contributions to robotic apple harvesting. First, the lightweight YOLOv8n–C2f–Faster model was applied for the first time to apple detection, demonstrating real-time perception capabilities in orchard environments. Second, a new path planning algorithm—DSA-BiTRRT—was proposed and successfully implemented in the context of dwarf hedge-planted orchards. Third, the findings emphasized the importance of considering the fruit weight after detachment when using soft grippers, as the failure to do so may result in dropped apples. The integration and field evaluation of the visual perception module, path planning algorithm, and adaptive soft gripper revealed a degree of effectiveness and operational practicality.

However, compared to existing robotic harvesting systems, there remains substantial room for improvement, particularly in reducing cycle times and increasing harvesting success rates. Prior studies have reported cycle times ranging from 4.0 to 17.17 s and success rates between 47.37% and 84.6% [5,21,22,48,49]. In this study, the proposed system achieved a harvesting success rate of 47.50% and an average cycle time of 15.26 s per apple. These results suggest the need for further enhancements in both efficiency and reliability. The low success rate was primarily attributed to camera localization errors and limited obstacle avoidance capabilities, while the extended cycle time was largely due to the long distance to the placement box, which required complex path planning and caused the placement phase to account for nearly half of the total operation time.

Building on these observed strengths and limitations in efficiency and reliability, the proposed system demonstrates several notable advantages, including a compact design, operational flexibility, and the seamless integration of perception and path planning modules. While the detachment methods explored in this study offer valuable preliminary insights into their feasibility and effectiveness, the sample size was relatively small. Although the tested samples provide an initial indication of the harvesting performance, the limited number of apples restricts the generalizability of the results. Future research should therefore include larger-scale trials across diverse orchards, canopy structures, and fruit maturity stages to improve the robustness and applicability of the findings. Additionally, potential interference from leaves positioned between the gripper and the fruit may cause occlusions, negatively impacting the detection and localization accuracy. Such interference can also reduce the grasp stability by increasing slippage or necessitating higher detachment forces. These factors should be systematically investigated in future studies. Moreover, as no post-harvest assessments were conducted to evaluate internal bruising or delayed physical damage, the effects on the fruit quality remain uncertain. Future work should incorporate extended storage evaluations to monitor both surface blemishes and the progression of internal damage over time.

To enhance the harvesting performance and reliability of the proposed system, future research should focus on the following key areas.

First, improving the detection accuracy is essential for overcoming positioning errors caused by lighting variations and occlusions, which hinder precise apple localization. Optimizing neural network architectures, incorporating multimodal data fusion—such as RGB, hyperspectral, and thermal imaging—and developing adaptive algorithms for dynamic parameter adjustments can significantly enhance the detection precision, reduce false identifications, and enhance the overall efficiency of the harvesting process. Furthermore, the limited number of camera viewpoints may restrict the accurate monitoring of fruit positions, especially in scenarios involving multiple robotic arms or where fruit displacement occurs during harvesting. To address this limitation, future studies should explore multi-view perception systems or implement real-time dynamic tracking techniques to support more complex and scalable harvesting operations.

Second, increasing the grasping precision and target segmentation is essential for minimizing mis-grasping, particularly in hedge-planted orchards where dense branches obstruct the robotic gripper. Integrating deep learning-based instance segmentation can enhance the differentiation between apples and branches. Additionally, incorporating vacuum suction systems alongside the adaptive gripper can enhance grasping stability, reduce failure rates, and enhance detachment reliability.

Third, real-time obstacle avoidance is essential for preventing collisions with branches and trellis structures, especially during the approach and placement. Integrating LiDAR, depth cameras, and fused imaging technologies can enhance perception and environmental awareness. Constructing 3D obstacle models to segment orchard structures, combined with collision-free trajectory planning, can enable adaptive motion control, improving operational efficiency and reliability in unstructured environments.

Fourth, optimizing the performance of the soft gripper is necessary for improving the handling and detachment. While the current gripper offers low-damage handling and mechanical flexibility, improvements in the load capacity and adaptability are needed to meet the demands of diverse orchard environments. Future research should explore the use of high-strength composite materials or variable-stiffness structures to enhance durability and optimize the force distribution during gripping. Additionally, it is important to systematically evaluate a broader and more granular range of torque levels under varying fruit maturities, stem characteristics, and canopy conditions to refine detachment strategies and minimize damage. The integration of force feedback sensors and real-time torque control would enable the gripper to dynamically adjust the pressure based on the fruit texture and size, reducing damage and improving detachment efficiency. Implementing fruit damage evaluation metrics will further ensure that the post-harvest quality aligns with commercial standards.

Finally, optimizing the placement strategy can significantly improve harvesting efficiency. The current method, which deposits apples at the rear of the vehicle, ensures a precise placement but prolongs the cycle time. Future enhancements should explore chute- or conveyor-based transfer mechanisms to streamline the placement while maintaining fruit integrity. Additionally, incorporating soft-padding or shock-absorbing structures can further reduce the impact damage, ensuring high-quality post-harvest handling.

4. Conclusions

This study presents a 6-DOF robotic arm system developed for harvesting apples in dwarf hedge-planted orchards. The system integrates the lightweight YOLOv8n–C2f–Faster algorithm for real-time apple detection and localization and the DSA-BiTRRT algorithm to enhance the path planning efficiency, trajectory quality, and operational success rates. A key advantage of the system relies on its ability to perform flexible picking and placement operations within the complex and constrained environments typical of orchard settings. Field experiments demonstrated that the direct-pull detachment method at 150 mN·m was the most effective, enabling a reliable apple detachment while providing adequate load-bearing support. In fully integrated harvesting trials, the system achieved an overall harvesting success rate of 47.50% and an average cycle time of 15.26 s per apple. To further enhance the system performance, future efforts should focus on improving obstacle avoidance and refining motion planning algorithms to reduce inefficiencies and failure rates. This study provides valuable insights into the current capabilities of autonomous harvesting systems and highlights key areas for future refinement and optimization.