Quantification and Optimization of Straight-Line Attitude Control for Orchard Weeding Robots Using Adaptive Pure Pursuit

Abstract

In automated orchard operations, the straight-line locomotion stability of ground-based weeding robots is critical for ensuring path coverage efficiency and operational reliability. To address the response lag and high-frequency oscillations often observed in conventional PID and fixed-lookahead Pure Pursuit controllers, this study proposes an adaptive lookahead Pure Pursuit method incorporating angular velocity feedback. By dynamically adjusting the lookahead distance according to real-time attitude changes, the method enhances coordination between path curvature and robot stability. To enable systematic evaluation, three time-series-based metrics are introduced: mean absolute yaw error (MAYE), peak-to-peak fluctuation amplitude, and the standard deviation of angular velocity, with overshoot occurrences included as an additional indicator. Field experiments demonstrate that the proposed method outperforms baseline algorithms, achieving lower yaw errors (0.61–0.66°), reduced maximum deviation (≤3.7°), and smaller steady-state variance (<0.44°²), thereby suppressing high-frequency jitter and improving turning convergence. Under typical working conditions, the method achieved a mean yaw deviation of 0.6602°, a fluctuation of 5.59°, an angular velocity standard deviation of 10.79°/s, and 155 overshoot instances. The yaw angle remained concentrated around the target orientation, while angular velocity responses stayed stable without loss-of-control events, indicating a favorable balance between responsiveness and smoothness. Overall, the study validates the robustness and adaptability of the proposed strategy in complex orchard scenarios and establishes a reusable evaluation framework, offering theoretical insights and practical guidance for intelligent agricultural machinery optimization.

1. Introduction

In orchard cultivation, weed growth not only competes with fruit trees for water, nutrients, and sunlight, but also serves as a potential host for pests and diseases [1], severely affecting tree development and fruit quality [2,3,4]. Therefore, weed management is a crucial aspect of orchard ground maintenance, playing a vital role in enhancing crop yields and ensuring ecological health. Traditional manual weeding is labor-intensive and inefficient, and with rising labor costs and an aging agricultural workforce, it can no longer meet the demand for efficient and precise orchard management [5]. With the rapid advancement of agricultural intelligent equipment, automated robots have been increasingly adopted for orchard operations [6,7,8,9,10]. Compared to manual methods, ground-based weeding robots offer advantages such as higher efficiency, reduced labor intensity, and consistent performance, making them a focal point of research in orchard ground management [11]. During inter-row weeding operations, the robot’s ability to maintain stable straight-line movement directly affects the completeness of path coverage and the overall operational stability [12,13], forming a foundational capability for high-quality task execution.

However, orchards represent a typical semi-structured environment characterized by narrow aisles, uneven terrain, and potential interference from tree branches or support structures [14,15,16,17]. As the robot navigates the terrain, it is susceptible to disturbances from the ground surface, perception errors, or minor deviations in its intended path. These factors often result in attitude drift or oscillations, manifested as significant yaw fluctuations and unstable angular velocity, which can lead to path discontinuities, trajectory overlap, or deviation from the intended route [18,19,20,21,22]. Therefore, enhancing the attitude control performance of weeding robots during straight-line motion is a key research direction for improving operational quality and automation levels.

Among current path control strategies for ground-based orchard robots, PID and Pure Pursuit are the most widely used. The former is a classic feedback control method, while the latter is a geometric path-tracking algorithm [23,24]. PID controllers are simple in structure and respond quickly, but are highly sensitive to parameter tuning and prone to attitude oscillations when facing path variations or external disturbances. In contrast, Pure Pursuit constructs a circular arc between the current position and a target point to achieve path tracking, offering intuitive geometric interpretation and high implementation efficiency, and is therefore commonly used in orchard path tracking tasks [25,26]. Previous studies have shown that although PID control can reduce trajectory deviation, it is prone to oscillations under complex conditions. In contrast, the Pure Pursuit algorithm, enhanced with error correction, has achieved sub-decimeter accuracy in field navigation [27]. Moreover, research on differential-drive robots has highlighted the critical influence of the look-ahead parameter in Pure Pursuit, demonstrating that proper tuning can effectively mitigate oscillations and improve tracking stability [28].

Nevertheless, both methods have inherent limitations in real-world orchard environments. Taking Pure Pursuit as an example, it relies on a fixed lookahead distance parameter, which is often inadequate under frequent disturbances and rapid attitude changes. A fixed lookahead distance cannot adapt to the robot’s dynamic response: during sharp turns, it fails to shorten the lookahead for fast correction; on relatively straight paths, it may cause over-adjustment, leading to unnecessary angular oscillations.

Moreover, most existing studies evaluate control performance using static indicators such as trajectory deviation or path error, with less emphasis on dynamic behavior during the attitude response process [29,30]. During straight-line motion, the robot’s attitude evolution exhibits clear temporal patterns—such as the gradual convergence of yaw angle and the rhythmic fluctuation of angular velocity—which reflect the controller’s responsiveness and stability. These time-series characteristics are critical for assessing control strategy performance. However, there is a lack of systematic quantitative metrics tailored to attitude response dynamics, and few comparative studies have explored the behavioral differences in various control methods under dynamic orchard scenarios.

To address these issues, this study proposes an adaptive lookahead Pure Pursuit control method based on angular velocity feedback. The proposed approach explicitly introduces the yaw rate ψ(t) as a modulation factor, establishing a ψ-driven coupling mechanism that directly links path geometry with attitude dynamics. When the robot experiences rapid fluctuations in orientation, the lookahead distance is shortened to improve steering responsiveness; under stable conditions, the lookahead distance is extended to suppress oscillations. This design achieves a favorable balance between responsiveness and stability. In addition, three time-series-based evaluation metrics—mean yaw error, maximum yaw deviation, and the standard deviation of angular velocity—are established to quantitatively characterize the robot’s attitude response during straight-line motion.

To validate the effectiveness of the proposed method, comparative experiments were conducted in a real pergola-style orchard. The robot was programmed to follow a “Z”-shaped bidirectional path, and three control strategies were tested under identical conditions: adaptive lookahead Pure Pursuit (with PID), original fixed-lookahead Pure Pursuit (with PID), and standalone PID without tracking assistance. Multiple sets of yaw and angular velocity data were collected and analyzed using time-series methods. The key contributions of this study are summarized as follows:

• An adaptive Pure Pursuit control strategy integrating yaw-rate feedback is proposed, further improving path-tracking accuracy and attitude stability.
• A set of time-series evaluation metrics tailored to attitude response is developed, providing a quantitative framework for straight-line control performance analysis.
• Comprehensive experiments in real orchard environments are conducted, systematically comparing attitude behaviors under different control strategies and demonstrating the effectiveness and application potential of the proposed method.

2. Materials and Methods

2.1. Experimental Data Acquisition

The straight-line motion and attitude data of the weeding robot used in this study were collected at the Yangyu Ecological Pear Orchard Base in Jiangsu Province, China (latitude 32.32° N, longitude 120.10° E, 2025). The test site adopts a standardized pergola-style fruit tree cultivation structure, which provides good inter-row visibility and reproducible experimental conditions. As illustrated in Figure 1, the orchard is constructed using a triangular frame layout, with a row spacing of 5 m, a plant spacing and trellis height of 3 m, and a frame inclination angle of approximately 70°. This structure not only ensures uniform light distribution for fruit tree growth, but also offers clear walking corridors and structural boundaries for robot path planning and attitude sensing.

During path control, the robot employed an RTK-GNSS system for initial localization, consisting of a portable base station (Rock-Smart A332 BD Base, China) and a rover receiver (Rock-Smart NZ500, China). The base station was deployed at the orchard edge to provide real-time differential corrections, enabling centimeter-level positioning accuracy. For attitude estimation, an onboard IMU (WHEELTEC H30, China) was mounted near the chassis center with its sensing axes aligned to the body frame. A static bias calibration was performed prior to experiments to minimize drift. The IMU integrates an accelerometer (±16 g), gyroscope (±2000°/s), and magnetometer (±49 Gauss), and its data were logged at 400 Hz and down-sampled to 100 Hz for control updates. GNSS position outputs and IMU attitude signals were further combined using a simple complementary filtering scheme, ensuring smooth and reliable state estimation. These specifications ensured reliable yaw-rate and posture feedback for the adaptive Pure Pursuit controller, which dynamically regulated the lookahead distance and enhanced straight-line attitude stability.

This study focuses specifically on the straight-line movement of the robot during orchard operations. Among various attitude parameters, yaw angle is selected as the primary variable for analysis. Compared to pitch and roll, the yaw angle more directly reflects horizontal path deviation and movement stability, making it a key indicator for evaluating straight-line tracking performance and control precision.

2.2. Quantitative Analysis and Optimization of Straight-Line Attitude Control

2.2.1. Pure Pursuit Control Algorithm

To achieve stable straight-line motion of the weeding robot within orchard corridors, this study adopts the Pure Pursuit algorithm as the baseline path-tracking control strategy. The Pure Pursuit algorithm guides the robot toward a designated lookahead point on the reference path by defining a lookahead distance and generating control commands based on geometric relationships. It features a simple structure, ease of implementation, and smooth response behavior. In semi-structured environments such as orchards, the robot must continuously navigate between regularly spaced rows while coping with challenges such as uneven terrain, minor path disturbances, and sensor noise. Owing to its real-time responsiveness and adaptability, Pure Pursuit can effectively track pre-defined paths while maintaining a degree of attitude stability—making it especially suitable for embedded systems with limited computational resources. As such, the algorithm serves as a solid foundational control framework in this study, providing a reliable baseline for introducing adaptive mechanisms and optimizing attitude control performance.

2.2.2. Adaptive Lookahead Pure Pursuit Control

To enhance the path-tracking stability and attitude regulation capability of the weeding robot during straight-line navigation in orchard corridors, this study proposes an improved adaptive lookahead Pure Pursuit control strategy based on angular velocity feedback. Concurrently, a quantitative analysis framework centered on yaw response characteristics is established to evaluate the control system’s attitude adjustment performance and stability during real-world operation. As illustrated in Figure 2, in the conventional Pure Pursuit algorithm, the lookahead distance L_d is predefined as a fixed constant, lacking the ability to adapt to attitude disturbances in dynamic environments. This often leads to delayed responses when the robot experiences minor path deviations or navigates segments with sudden curvature changes.

To improve the controller’s responsiveness to attitude variations, the yaw rate of the robot ψ(t) is introduced as a modulation factor. Based on this, an adaptive mechanism is designed to update the lookahead distance dynamically. The proposed update rule is formulated as follows:

$$L_d(t)=\mathrm{clip}\left(L_{\mathit{0}}-k\cdot \left|\psi (t)\right|;L_{\mathit{min}},L_{\mathit{max}}\right)$$

(1)

$$\begin{array}{l}\delta (t)=sat\left(\arctan {\displaystyle \frac{2L\cdot \sin (\alpha )}{L_d(t)}};-{\delta }_{\mathit{max}},{\delta }_{\mathit{max}}\right)\\ \end{array}$$

(2)

In the equations, L_d(t) denotes the dynamic lookahead distance (m) that adapts in real time to attitude responses within the bounds [L_min, L_max]. L₀ is the baseline lookahead distance (m), k is the angular velocity sensitivity gain with units of m·s/deg (yaw-rate is in °/s), and ψ(t) = Δyaw/Δt refers to the yaw rate (°/s). The operator clip(·) enforces clamping between L_min and L_max. The variable δ(t) is the steering command, limited by the saturation function sat(·) to the actuator bounds [−δ_max, δ_max]. L is the vehicle wheelbase (m), and α represents the yaw angle difference (°) between the lookahead point and the vehicle centroid line.

This formulation preserves the geometric simplicity of the original Pure Pursuit algorithm while embedding a dynamic coupling mechanism that explicitly links path curvature with instantaneous attitude response. When the robot undergoes rapid fluctuations in orientation, the lookahead distance is shortened to enhance steering responsiveness and improve the timeliness of trajectory correction. Conversely, under stable conditions, the lookahead distance is extended to mitigate control oscillations and promote smoother trajectory following. This adaptive mechanism balances tracking accuracy with motion stability, thereby enhancing the robustness of autonomous navigation in complex environments. To ensure the reproducibility of the results, the main control parameters—including the baseline lookahead distance, coupling gain, and clamping limits—are specified, as summarized in

Table 1. Parameterization of the adaptive lookahead Pure Pursuit controller.

Parameter	Symbol	Value	Unit
Baseline lookahead distance	L0	1.00	m
Coupling gain	k	0.25	m·s/°
Minimum lookahead distance	Lmin	0.60	m
Maximum lookahead distance	Lmax	1.60	m

.

2.2.3. Quantitative Analysis of Straight-Line Attitude Response

To comprehensively evaluate the regulation performance of the improved control strategy during actual operation, this study constructs a set of time-series quantitative indicators based on the robot’s yaw angle sequence yaw(t) and angular velocity sequence ψ(t) [31]. These indicators are designed to characterize the stability and adjustment trends of the control response process, thereby forming a unified closed-loop framework of “control—response—evaluation”.

Three key indicators are introduced. The mean yaw angle deviation reflects the overall offset of the robot’s orientation from the desired trajectory, serving as a measure of directional stability. The peak-to-peak amplitude captures the maximum oscillation range of the yaw angle, enabling the detection of abrupt attitude fluctuations. The standard deviation of angular velocity quantifies the variability of the yaw rate over time and is used to assess the degree of jitter during regulation. Together, these three metrics establish a robust quantitative framework for evaluating the effectiveness of the control strategy in maintaining stable attitude regulation during straight-line motion. Their mathematical definitions are given as follows:

$$E\phi ={\displaystyle \frac{1}{T}}{\displaystyle \sum _{t=1}^T\left|\phi (t)\right|}$$

(3)

$$A_{pp}=\max \left(\phi (t)\right)-\min \left(\phi (t)\right)$$

(4)

$${\sigma }_{\psi }=\sqrt{{\displaystyle \frac{1}{T}}\sum _{t=1}^T{\left(\psi \left(t\right)-\theta \right)}^2}$$

(5)

In the equations, E_φ denotes the average yaw angle deviation, T is the total number of sampled frames, and φ(t) represents the yaw angle deviation of the robot at time t, reflecting whether the robot continuously deviates from the target heading during the entire operation period; smaller values indicate a more stable navigation posture. A_pp is the peak-to-peak amplitude of the yaw angle signal within a specific operation segment, characterizing the extreme variations in the robot’s posture response during that period. σ_ψ represents the standard deviation of the yaw rate, where ψ(t) is the yaw rate at time t and θ is its mean value, reflecting the dispersion of adjustment amplitudes during posture regulation; smaller values indicate a smoother response.

Finally, to characterize the nonlinear oscillatory behavior during the attitude adjustment process, the overshoot count is introduced as a supplementary criterion for evaluating dynamic regulation trends. Specifically, an overshoot is defined as a yaw angle deviation exceeding a threshold of ±0.5° at a local extremum. No hysteresis band was applied in order to ensure comprehensive quantification, allowing all deviations beyond the threshold to be captured. The overshoot count thus quantifies the frequency of meaningful over-adjustments under external disturbances, reflecting the extent of frequent attitude corrections. In summary, the proposed adaptive control strategy enhances the responsiveness of the robot to path disturbances by dynamically adjusting the lookahead distance, while the multidimensional set of quantitative indicators provides a systematic framework for evaluating attitude regulation performance in dynamic orchard environments.

3. Results and Discussion

3.1. Analysis of Straight-Line Attitude Control Performance

To validate the performance advantages of the proposed adaptive lookahead Pure Pursuit control algorithm in straight-line motion tasks, a comparative experiment was conducted involving three control strategies: (i) the improved adaptive lookahead Pure Pursuit algorithm (combined with PID), (ii) the original fixed-lookahead Pure Pursuit algorithm (also combined with PID), and (iii) a standalone PID controller without path tracking. The comparative analysis focuses on evaluating the stability and control accuracy of each method during actual operation. Three performance metrics were introduced to comprehensively assess control effectiveness: mean absolute yaw error, maximum yaw deviation, and steady-state attitude variance. These indicators, respectively, reflect overall control accuracy, response to extreme deviations, and the controller’s ability to maintain attitude stability and quick convergence. The aim is to objectively evaluate each control method’s performance under identical environmental conditions in terms of straight-line tracking capability.

The experiment was conducted within orchard operation corridors, where the robot continuously traversed a predefined “Z”-shaped bidirectional path at a constant forward velocity of 0.5 m/s to ensure full coverage of the orchard area. The onboard IMU (H30, WHEELTEC) was employed to continuously and synchronously collect the real-time yaw angle of the weeding robot, which served as the primary variable for evaluating straight-line attitude control performance. In addition to yaw, the IMU also outputs nine-axis sensing data (acceleration, angular velocity, magnetometer readings, and Euler angles), but yaw was specifically selected as the core indicator in this study due to its direct relationship with lateral path deviation. The IMU data were transmitted via UART to the Jetson AGX Orin in binary frame format at 400 Hz, ensuring high-frequency and low-latency acquisition for dynamic analysis. For control, the yaw-rate signals were further low-pass filtered and down-sampled to a 100 Hz update rate.

To ensure reproducibility, the comparator controllers were fully specified. The PID controller was applied to the heading error (yaw angle deviation) with gains set to K_p = 0.8, K_i = 0.05, and K_d = 0.1; an integral limit was included to avoid windup, and actuator commands were saturated at ±δ_max according to the motor driver limits. For the fixed-lookahead Pure Pursuit controller, the lookahead distance was set to L₀ = 1.0 m, consistent with the chassis scale, and combined with the same PID stabilization. All control algorithms were deployed on the same edge computing platform—Jetson AGX Orin (featuring a 12-core Arm Cortex-A78AE v8.2 64-bit CPU with 3 MB L2 and 6 MB L3 cache, 2048 CUDA cores, 64 Tensor Cores, and 64 GB LPDDR5 memory; MAKEROBOT, China). Control commands generated by the algorithm were further transmitted via a CAN bus interface to the motor drivers, forming a complete closed-loop data transmission architecture. In each trial, a representative segment containing 3500 yaw data frames was randomly selected from the complete trajectory for subsequent analysis. An example is shown in Figure 3.

As shown in Figure 3, the three control algorithms exhibit distinct differences in the yaw response characteristics of the weeding robot. In Figure 3a, the adaptive lookahead Pure Pursuit controller maintains a consistently low amplitude of attitude fluctuation throughout the entire operation segment. The response curve remains smooth without significant overshoot, indicating strong disturbance rejection capability and overall control stability. In contrast, while the original Pure Pursuit algorithm retains basic control accuracy over most of the trajectory, it shows signs of response delay during segments with intensified path disturbances, suggesting insufficient sensitivity in its regulation process. The PID controller, on the other hand, displays pronounced high-frequency oscillations across the entire segment, leading to unstable attitude behavior and an increased risk of path deviation.

A closer examination of the two locally magnified time segments in Figure 3b,c further highlights these differences. The adaptive controller demonstrates rapid suppression of attitude deviations during disturbance events, with fast convergence and superior smoothness compared to the other two methods. The response curve of the original Pure Pursuit controller exhibits larger fluctuations and moderate drifting tendencies, while the PID controller shows strong periodic oscillations in both segments. Such high-frequency responses could seriously compromise path consistency and operational precision during orchard tasks.

In summary, by incorporating angular velocity feedback and dynamically adjusting the lookahead distance, the adaptive Pure Pursuit controller significantly improves the responsiveness and robustness of path tracking. Compared to traditional control methods, it offers distinct advantages in handling complex trajectories and unstructured environmental disturbances common in orchards, demonstrating promising potential for practical engineering applications.

To ensure the reliability and statistical representativeness of the experimental results, each of the three control methods was independently repeated five times, and multiple statistical analyses were conducted based on the recorded results from each trial.

Table 2. Yaw attitude performance comparison of straight-line control algorithms.

Trial No.	Control Algorithm	Mean Absolute Yaw Error (°)	Maximum Yaw Deviation (°)	Steady-State Attitude Difference (°2)
1	Adaptive Pure Pursuit	0.66	3.69	0.43488
	Pure Pursuit	1.25	3.67	0.47084
	PID controller	1.53	4.15	0.52011
2	Adaptive Pure Pursuit	0.62	3.55	0.42831
	Pure Pursuit	1.19	3.60	0.46592
	PID controller	1.50	4.10	0.51235
3	Adaptive Pure Pursuit	0.64	3.62	0.43214
	Pure Pursuit	1.23	3.65	0.46918
	PID controller	1.56	4.20	0.52843
4	Adaptive Pure Pursuit	0.61	3.50	0.42567
	Pure Pursuit	1.20	3.55	0.46325
	PID controller	1.52	4.12	0.51794
5	Adaptive Pure Pursuit	0.65	3.68	0.43129
	Pure Pursuit	1.24	3.66	0.47121
	PID controller	1.54	4.18	0.52316

presents the control performance of the three algorithms on the same segment.

Based on the data presented in Table 2, the adaptive lookahead Pure Pursuit controller demonstrates superior performance across all metrics. The mean absolute yaw error is consistently controlled between 0.61° and 0.66°, with the maximum yaw deviation not exceeding 3.7° and steady-state variance fluctuations below 0.44°². These results indicate that the method offers strong control accuracy and attitude stability during straight-line motion. In comparison, the original Pure Pursuit controller shows good overall deviation control but exhibits slightly higher steady-state variance, suggesting some residual oscillation. On the other hand, the traditional PID controller consistently exhibits large yaw errors and significant fluctuations across all trials, with maximum yaw deviations generally exceeding 4.1° and steady-state variance considerably higher than the other two methods. This suggests that PID control is more prone to high-frequency oscillations during path disturbances or attitude corrections, negatively impacting overall control stability.

To further quantify the consistency and statistical significance of the experimental outcomes, the results of five independent trials for each control method were summarized using descriptive statistics, as shown in

Table 3. Statistical summary of three straight-line control algorithms (Mean ± Std. Dev. with 95% Confidence Intervals).

Algorithm	Mean Absolute Yaw Error (°)	Maximum Yaw Deviation (°)	Steady-State Attitude Difference (°2)
Adaptive Pure Pursuit	0.636 ± 0.021 (95% CI: 0.61–0.66)	3.608 ± 0.082 (95% CI: 3.51–3.71)	0.430 ± 0.004 (95% CI: 0.426–0.435)
Pure Pursuit	1.222 ± 0.026 (95% CI: 1.19–1.26)	3.626 ± 0.050 (95% CI: 3.56–3.69)	0.468 ± 0.003 (95% CI: 0.464–0.472)
PID Controller	1.529 ± 0.022 (95% CI: 1.50–1.56)	4.151 ± 0.041 (95% CI: 4.10–4.20)	0.520 ± 0.006 (95% CI: 0.513–0.528)
Adaptive Pure Pursuit	0.636 ± 0.021 (95% CI: 0.61–0.66)	3.608 ± 0.082 (95% CI: 3.51–3.71)	0.430 ± 0.004 (95% CI: 0.426–0.435)

.

As shown in Table 3, the adaptive lookahead Pure Pursuit controller consistently achieves the lowest mean absolute yaw error (0.636°) and the smallest steady-state variance (0.430°²), with the narrowest 95% confidence intervals across all metrics. This indicates that the method not only improves control accuracy but also enhances reliability by reducing variability across repeated trials. In contrast, the original Pure Pursuit algorithm achieves moderate performance, with slightly higher yaw errors and variance, suggesting residual oscillatory behavior. The PID controller performs the worst among the three, exhibiting significantly larger yaw errors (1.53° on average) and higher steady-state variance (0.520°²), reflecting its susceptibility to high-frequency oscillations. These findings confirm that the adaptive Pure Pursuit algorithm offers both superior accuracy and greater robustness in orchard straight-line navigation tasks.

In conclusion, the improved adaptive lookahead Pure Pursuit control algorithm demonstrates stronger robustness and consistency in straight-line path tracking tasks within orchard environments, highlighting its greater potential for practical application and widespread adoption.

3.2. Time-Series Behavior Analysis of Attitude Response Characteristics

To further assess the regulation performance of the proposed control strategy during actual operation, this study collects yaw angle sequences yaw(t) and angular velocity sequences ψ(t) during robot operation. The analysis introduces several key metrics: mean yaw deviation, peak-to-peak fluctuation amplitude, angular velocity standard deviation, and overshoot count. These metrics are employed to provide a deeper insight into the stability and regulation trends during the control response process, forming an integrated “control—response—evaluation” closed-loop framework. The statistical results are shown in Figure 4.

The visualization results in Figure 4 further reveal the dynamic response characteristics and regulation trends of the control strategy during actual operation. Figure 4a shows the instantaneous angular velocity variation curve of the weeding robot over the full operational segment. The overall fluctuation range is relatively small, with no significant jumps or sustained instability, indicating that the attitude adjustment process is relatively smooth. To avoid interference from extreme values, an anomalous deviation point (ψ = 396°/s) has been excluded, ensuring that the statistical features are more representative. Figure 4b presents the kernel density estimate (KDE) distribution of the yaw angle, which shows a slight right-skewed characteristic. The main density is concentrated in the range of −1° to 0.5°, indicating that for most of the time, the robot’s yaw angle remains within a small deviation range, reflecting good attitude stability. Figure 4c plots the complete yaw angle variation curve based on the time-series, with 155 overshoot points (red dots) and all extreme points (red crosses) marked, reflecting the system’s ability to identify and adjust deviations in real-time during the dynamic process.

Specifically, the mean yaw deviation is 0.6602°, indicating a small overall attitude offset and good path-holding ability. The peak-to-peak fluctuation amplitude reaches 5.5900°, suggesting that transient large disturbances occur during some periods, but the overall fluctuation distribution is relatively dispersed, without a systemic loss of control. The standard deviation of angular velocity is 10.79°/s, indicating that the fluctuation amplitude during the adjustment process is moderate, with no significant oscillation. In addition, a total of 155 overshoot points were detected, primarily concentrated in segments with sharp attitude changes, showing that the system has a strong response ability to minor disturbances and can quickly correct yaw errors to maintain overall operational stability.

In conclusion, the proposed adaptive control strategy demonstrates excellent response speed and fluctuation control capability during dynamic regulation. It can quickly correct deviations under disturbances and suppress overshoot, effectively balancing stability and sensitivity. The “angle–velocity–overshoot” multidimensional quantitative indicator system developed in this study provides a universal analysis framework for the time-series evaluation and comparison of control strategies, with significant application value and potential for use in complex orchard path conditions.

4. Conclusions

This study proposed an adaptive lookahead Pure Pursuit control method based on angular velocity feedback to enhance the attitude stability and tracking accuracy of orchard weeding robots under disturbed path conditions. By dynamically adjusting the lookahead distance according to yaw rate changes, the method realized coupled regulation of path curvature and attitude response within a unified “control–response–evaluation” framework.

Field experiments on a predefined “Z”-shaped orchard path demonstrated that the proposed method consistently outperformed the original Pure Pursuit and PID controllers. It achieved mean absolute yaw errors of 0.61–0.66°, maximum yaw deviations below 3.7°, and steady-state variances under 0.44°². These results confirm the ability of the method to suppress high-frequency oscillations and accelerate attitude convergence, thereby improving robustness and stability.

Time-series analysis of yaw angle and angular velocity further validated its performance, yielding a yaw deviation of 0.6602°, fluctuation amplitude of 5.5900°, and angular velocity standard deviation of 10.79°/s, with 155 overshoot events. The occurrence of overshoot events is crucial as it reflects the system’s ability to recover from sharp deviations and maintain stability during transient disturbances. In real-world scenarios, this could indicate the effectiveness of the control method in reducing oscillations and improving the reliability of the robot’s navigation. Kernel density estimation showed yaw angles concentrated between −1° and 0.5°, while angular velocity remained stable, highlighting balanced responsiveness and stability.

The proposed control method demonstrates strong potential for orchard navigation tasks and provides a transferable analytical framework for evaluating agricultural robot control systems. Nevertheless, this study is limited by experiments conducted in controlled orchard corridors and single-robot scenarios, which may restrict its generalization to more complex and unstructured environments. The method was tested only on flat terrain and straight-line paths, which may not fully capture the challenges of curved paths or sloped orchards. Future work will extend the method to curved paths, variable payloads, and multi-robot cooperation, while integrating predictive and learning-based strategies to further enhance robustness and adaptability in precision agriculture. Additionally, we will explore the method’s applicability to multi-row operations and turning maneuvers in orchard navigation.