This study aims to develop an autonomous driving algorithm simulation framework for small unmanned harvesters. However, due to the lack of a high-fidelity specialized model precisely matching the target platform—a small full-feed tracked combine harvester—in the standard model library of the selected simulation software (TruckSim), this study strategically selected the built-in TruckSim model (used for theoretical model derivation in Section 1 of this chapter) and the Suzhou King Long KLQ6125D bus model (used for vehicle dynamics simulation and control system validation in Section 4 of this chapter) as alternative simulation platforms (Mechanical Simulation Corporation, 2021). This choice is based on the following considerations:

In the established dynamic model, the dynamic behavior of the tractor and semi-trailer is significantly influenced by the road adhesion coefficient μ. μ not only determines the vehicle’s traction force and stability limits but also directly affects path planning and control system performance. Therefore, to achieve high-performance autonomous driving, this paper employs the Extended Kalman Filter (EKF) method to identify the road adhesion coefficient in real time (Ding et al., 2022). EKF utilizes the system state equations, observation equations, and the statistical characteristics of process noise and measurement noise to perform optimal estimation, enabling accurate estimation of μ and providing critical road condition information to the control system (Bavdekar et al., 2011).

Real-time identification of μ is crucial for dynamically adjusting vehicle control strategies, especially in complex and variable farmland road conditions. For example, under low μ conditions, vehicle handling stability decreases, and the system must incorporate the real-time identified μ results to ensure driving safety through path planning (such as the hybrid A* algorithm) and control decisions (such as PID parameter tuning or speed limiting).

To balance computational complexity and algorithm feasibility verification, a nonlinear three-degree-of-freedom vehicle model is used as the basis for the EKF. The core design of the EKF relies on four types of information: state equations, observation equations, process noise statistical characteristics, and measurement noise statistical characteristics. EKF achieves local linearization of the nonlinear model by performing a first-order Taylor expansion of the nonlinear functions at the current state estimate point (neglecting higher-order terms), making it widely applicable to various nonlinear system estimation problems. The EKF-based road adhesion coefficient identification process includes the following steps:

In autonomous driving systems, path planning and vehicle control are tightly coupled critical components. The Hybrid A* algorithm combines the efficient search capabilities of the traditional A* algorithm with the advantages of other methods (such as considering vehicle dynamic constraints), enabling the generation of feasible paths that better align with the vehicle’s motion characteristics (Dolgov et al., 2010). The standard A* algorithm uses heuristic search to find the shortest path in a discrete graph but is less efficient in high-dimensional continuous state spaces or environments with complex constraints (Hart et al., 1968). By integrating multiple heuristic rules and directly sampling in continuous state space, Hybrid A* can provide better solutions in dynamic environments and effectively incorporate real-time road adhesion coefficient μ information for risk assessment and path adjustment.

When μ changes, the set of feasible paths and safety margins for the vehicle also change. Under low μ conditions, planned paths need to avoid sharp turns and high-curvature sections while maintaining larger safety distances. The Hybrid A* algorithm not only needs to optimize traditional metrics such as path length or smoothness but also must evaluate path feasibility and risk by integrating real-time μ information, dynamically adjusting planning strategies—for example, avoiding low-μ areas or selecting more conservative routes. The flowchart of the Hybrid A* algorithm is shown in Figure 6.

Hybrid A* can be regarded as a path planning algorithm extended from the A* framework. Its core innovation lies in the use of dual heuristic functions: one is the cost function h1(x), and the other is the heuristic function h2(x). h1(x) represents the minimum cost from the current state x to the goal state (usually considering non-holonomic constraints, dynamic constraints, etc., which makes the computation complex), while h2(x) represents an optimistic estimated cost from the current state x to the goal state (typically using simplified models such as Reeds-Shepp curves, Euclidean distance, etc., which are computationally efficient). By combining h1(x) and h2(x), the Hybrid A* algorithm can significantly reduce the search time of the A* algorithm while ensuring path quality. The specific implementation method is as follows:

To quantitatively evaluate the performance advantages of the hybrid A* algorithm, this study conducted systematic comparative testing between the hybrid A* and traditional A* algorithms within identical simulation environments. The test scenario comprised a typical agricultural field area (dimensions 60 m × 40 m) containing six static obstacles, with the start and end points situated at opposite corners of the area. The performance comparison results are as follows:

1. Search efficiency improved by 40%

The traditional A* algorithm requires an average expansion of approximately 1,850 nodes, with an average computation time of 2.8 seconds; the hybrid A* algorithm requires an average expansion of approximately 1,120 nodes, with an average computation time of 1.68 seconds. The computational efficiency is as shown in Equation 15.

Among them, TA∗ and THybridA∗ denote the average single-run planning time for the two algorithms respectively. The efficiency gain primarily stems from Hybrid A* employing dual heuristic functions h1(x) and h2(x), which converge earlier to feasible paths during node expansion, thereby reducing futile searches.

2. Obstacle avoidance path length increased by only 12%

The average path length planned by the traditional A* algorithm was 52.4 metres; the hybrid A* algorithm, designed to accommodate vehicle kinematic constraints and achieve smooth steering, yielded an average path length of 58.7 meters. The proportional increase in path length is shown in Equation 16.

This increase in length is primarily attributable to the hybrid A* algorithm, which incorporates steering continuity constraints and path smoothing. This approach mitigates the abrupt turns and path jitter inherent in traditional A*, better aligning with the actual motion characteristics of vehicles. Consequently, while ensuring path feasibility, it achieves a modest extension of the travel distance.

3. Replanning response time< 0.1 seconds

In dynamic obstacle testing scenarios (simulating sudden moving obstacles encountered in the field), the system triggers its replanning mechanism. Simulation records indicate that the average time from obstacle detection to completion of new path generation is 0.086 seconds, with a standard deviation of 0.015 seconds, meeting real-time requirements (<0.1 seconds). This performance is primarily attributable to the hybrid A* algorithm’s mechanism for locally refining existing paths, coupled with the effective guidance provided by the heuristic function.

Finally, comparisons were conducted with three typical planning algorithms: the traditional A* algorithm, Dijkstra’s algorithm, and the Rapidly Exploring Random Tree (RRT) algorithm. Comparison metrics included planning time, path length, path smoothness (mean curvature), and number of nodes searched. The results are presented in Table 2.

Comparative analysis indicates that Hybrid A* achieves a 40% improvement in planning time over traditional A*, and approximately 60% over Dijkstra’s algorithm. Whilst slightly slower than RRT, it delivers markedly superior path quality. Regarding path length, Hybrid A* increases by only 12% compared to traditional A*, and is significantly shorter than RRT (62.3 m). Path smoothness (mean curvature) surpasses both traditional A* and Dijkstra’s algorithms, approaching RRT levels. The number of searched nodes is reduced by approximately 40% compared to traditional A*, demonstrating the efficiency advantage of dual heuristics.

In summary, Hybrid A* achieves a favorable balance between planning efficiency, path quality, and driving smoothness. It is particularly well-suited for operational tasks in agricultural settings demanding high real-time performance, safety, and driving stability.

Based on the above content, this paper uses the hybrid A* algorithm for path planning to enable the unmanned harvester to automatically find the optimal path during operation. As shown in Figure 7a, the obstacles are marked with asterisks, with the red asterisks indicating the start and end points. The nodes explored by the algorithm during the search for the optimal path are connected by red lines. Figure 7b shows the optimal path planned by the hybrid A* algorithm. Simulation results demonstrate that the hybrid A* algorithm successfully plans an optimal (or feasible) path that avoids all obstacles. This result validates the effectiveness of the hybrid A* algorithm in complex static farmland environments. More importantly, the algorithm framework has the potential to integrate real-time road adhesion coefficient μ information, allowing dynamic adjustment of the path cost function based on road conditions to select safer and more feasible paths. This enhances the system’s adaptability and safety in dynamic farmland environments, showcasing its application potential in the field of agricultural automation.

Whole vehicle modeling is a key step in vehicle dynamic simulation, primarily involving the modeling of sprung mass and unsprung mass. Sprung mass modeling includes parameters such as the vehicle body dimensions, center of gravity location, and moments of inertia, all of which significantly affect the vehicle’s stability and handling. Unsprung mass focuses on components not directly supported by the suspension system, such as the wheels, drive shafts, and braking system, and is related to parameters like the front and rear track widths and the wheels’ moments of inertia. The virtual vehicle prototype used in this simulation is the Suzhou King Long KLQ6125D series bus, with a curb weight of 5,025 kg, to align with the actual test vehicle, the mass and inertia parameters of this model have been scaled and adjusted in accordance with the equivalence principle outlined in Section 2.1, thereby simulating the target harvester’s dynamic characteristics, of which 4,455 kg is sprung mass and 570 kg is unsprung mass. The unsprung mass and parameters such as the front and rear track widths will be integrated with the suspension system modeling to ensure that the dynamic simulation results align with the actual vehicle performance.

Tires are an important component in vehicle dynamics analysis, as tire characteristics directly affect a vehicle’s power performance, braking, handling stability, ride comfort, and safety. In TruckSim, tire modeling mainly includes the tire’s geometric dimensions, steady-state mechanical characteristics, transient response characteristics, and dynamic hysteresis losses (Svendenius, 2007). Among these, the most critical aspect is the modeling of tire mechanical characteristics. This part of the modeling can be based on experimental data or can use empirical or semi-empirical tire models through parameter settings. Tire mechanical characteristics include longitudinal force, lateral force, and self-aligning torque. Accurate tire characteristics are essential to correctly reflect the vehicle’s state during simulation. This paper uses the Magic Formula tire model to model this tire type, as shown in Figure 8. This modeling approach is simple and can represent the tire’s mechanical characteristics both under pure conditions and combined conditions (Pacejka, 2012).

Further analysis of the rationality is presented in Figure 9. Figure 9a shows the longitudinal force characteristics of the tire, which are the most comparable equivalent characteristics to those of the track. The track system also exhibits a “slip ratio - traction force” relationship curve with a similar shape. Figure 9b illustrates the longitudinal force under combined slip; this 3D surface plot demonstrates how the longitudinal force diminishes when both longitudinal slip and slip angle are present simultaneously. In other words, during acceleration while turning, the effectiveness of the driving force is affected, which conceptually can be considered equivalent to the track characteristics.

The road environment settings are used to simulate common driving scenarios for vehicles, including straight roads, lane changes, curves, uphill climbs, and complex conditions such as slippery surfaces. In TruckSim, three-dimensional road surfaces can be created, allowing users to freely set road elevation, lateral and longitudinal slopes, and customize road shapes like sharp curves and gentle bends to realistically simulate various road environments. Additionally, TruckSim allows adjustment of the road adhesion coefficient to simulate hazardous conditions with low adhesion, especially under extreme road conditions such as icy or wet surfaces. In the road settings, it is also possible to create two-way roads or joint road surfaces for analyzing vehicle braking performance.

Other systems, such as the steering system and suspension system, will not be elaborated on in this text.

The output of path planning is the desired path (or trajectory), while lane-keeping control is the key execution layer for achieving precise path tracking. The lane-keeping system proposed in this paper combines a PID controller with a hybrid A* planned path to ensure stable operation of an unmanned harvester in complex agricultural environments. The PID controller makes real-time adjustments based on the lateral deviation (e) between the vehicle’s current position and the desired path (the lane center reference line), outputting front wheel steering angle commands ( δ_cmd) to enable the vehicle to follow the reference path (Ogata, 2010). This control process dynamically incorporates the real-time identified road adhesion coefficient μ, for example, increasing control margins or limiting maximum steering angle/acceleration under low μ conditions.

In this paper, the function of the PID controller is to calculate a new steering adjustment by using the vehicle’s lateral deviation as a feedback signal. Specifically, the PID controller continuously computes the lateral deviation and adjusts the steering angle magnitude to gradually reduce the vehicle’s distance from the lane. When the vehicle deviates significantly from the lane, the PID controller increases the steering angle to quickly guide the vehicle back to the lane center; when the vehicle is close to the lane center, the PID controller reduces the steering adjustment to avoid overcorrection. The feedback loop is the core part of this system—the feedback signal (i.e., lateral distance) is continuously fed back to the PID controller, and the system ensures the vehicle remains centered in the lane by real-time steering adjustments. The PID parameters of the controller were determined through a systematic process: first, baseline parameters were obtained using the Ziegler-Nichols method based on the vehicle lateral dynamics model; subsequently, within the TruckSim/Simulink co-simulation environment, fine-tuning was performed for typical farmland scenarios (straight lines, bends) via trial-and-error, targeting minimization of the integral absolute error; Finally, parameter sensitivity analysis (Figure 10) validated the optimality of the selected parameter combination (Kp=1.2, Ki=0.05, Kd=0.12) under the specified operating conditions.

In addition, the system may also use some mathematical transformations (such as the sine and cosine functions of trigonometry) to describe the geometric relationship between the vehicle’s position and steering angle, making the PID controller’s response more precise and efficient. Ultimately, the PID controller finely adjusts the steering angle, allowing the vehicle to smoothly stay within the lane and avoid drifting out of it.

The performance of the PID controller is highly dependent on parameter tuning. The sensitivity analysis in Figure 10 reveals key patterns: 1) When the proportional gain Kp = 1.2 and the derivative gain Kd = 0.12 (marked by the red point), the control performance reaches its peak (deep yellow region); 2) Changes in the derivative gain Kd have a more significant impact on system stability (contour lines are denser), indicating that during real vehicle debugging, priority should be given to ensuring the accuracy of the derivative term. Based on this analysis, this study sets the PID controller parameters to Kp = 1.2, Ki = 0.05, and Kd = 0.12, laying the foundation for subsequent lane-keeping control.

In the Simulink environment, a complete PID controller model can be constructed. The model’s input is typically the vehicle’s lateral deviation, while the output is the steering command δ_cmd. To better simulate real driving conditions, the model also needs to incorporate the vehicle’s dynamic characteristics, including the position of the vehicle’s center of gravity, moment of inertia, and tire properties. To further validate and optimize the PID controller’s performance, this study integrates simulation with TruckSim software. TruckSim can simulate vehicle behavior under real road conditions, receiving control signals output from Simulink and providing feedback on the vehicle’s actual driving state. By coordinating these two tools, the effectiveness of the PID controller can be efficiently tested in a virtual environment, avoiding potential safety risks associated with real-world road testing.

The vehicle model parameters used in the simulation process are consistent with those of the actual vehicle. The simulation model is shown in Figure 11. Other simulation parameters are listed in Table 3.

Figure 12 intuitively demonstrates the core performance of the PID controller in path tracking. Over the 120-second simulation period, the actual driving trajectory (red solid line) closely follows the reference path (blue dashed line). Even in the complex S-curve section between t=30 and 50 seconds, the maximum lateral deviation of the actual path is only 0.12 meters, consistently maintained within the ±0.15 meter safety margin (light blue shaded area). During the curved section (approximately X = 25–35 m), the vehicle trajectory shows slight overshoot but quickly converges back to the reference path, demonstrating the controller’s robustness.

Figure 13 shows the trajectory of the lane-keeping system, indicating that the vehicle adjusts its path through a PID controller to stay centered in the lane. In practical applications, the PID controller makes real-time adjustments based on the vehicle’s angle or distance of deviation from the lane, ensuring lateral stability. The paths in the figure demonstrate how the vehicle drives smoothly under different conditions. The trajectory presents a relatively stable driving path, indicating that the system can effectively adjust the vehicle’s direction to prevent it from drifting out of the lane.

Figure 14 shows the vehicle’s speed. The speed curve exhibits minor fluctuations, indicating that the vehicle maintains a relatively stable speed during path adjustments. This is key to the PID controller ensuring smooth driving while keeping the vehicle within the lane. Although there are some speed variations, they remain within an acceptable range, demonstrating that the vehicle’s lateral stability and longitudinal control work in coordination. The coordination between speed and path planning ensures the vehicle’s adaptability under different operating conditions.

Figure 15 shows the changes in the vehicle’s steering angle, which fluctuate quite dramatically. These fluctuations are caused by the PID controller adjusting the steering angle to minimize the vehicle’s deviation from the center of the lane. Specifically, when the vehicle is far from the lane center, the PID controller increases the steering angle to quickly guide the vehicle back to the lane center; when the vehicle is close to the lane center, the PID controller reduces the amount of steering adjustment to avoid overcorrection. The changes in the steering angle reflect the PID controller’s fine-tuned adjustments to the vehicle’s behavior.

By combining lane-keeping trajectories with vehicle speed and steering angle data, the effective application of the PID controller in the lane-keeping system can be observed. Fluctuations in speed and steering angle indicate that the PID controller can quickly respond to the vehicle drifting out of the lane and maintain the vehicle within the lane by adjusting the steering angle. Although the steering adjustments are relatively frequent, this also demonstrates the control system’s efficiency and flexibility in handling dynamic environments and complex paths. The lane-keeping trajectory shown in the image illustrates how the PID controller keeps the vehicle stable under various conditions while minimizing deviation to the greatest extent. For future research, PID controller parameters can be further optimized to explore how to improve the system’s robustness and adaptability in more complex environmental conditions. Additionally, integrating other control methods, such as Model Predictive Control (MPC) or Deep Reinforcement Learning (DRL), could be considered to further enhance the system’s performance and stability.

Further analyzing the control errors, Figure 16 reveals the system’s performance characteristics under different road conditions: 1) On dry pavement (green background, 0-40s), the lateral error remains stable below 0.05m, and the heading error is less than 1°; 2) On wet pavement (blue background, 40-80s), the lateral error slightly increases to 0.08m, with a peak heading error of 1.8°; 3) On muddy pavement (brown background, 80-120s), the control challenge is greatest, with lateral errors reaching up to 0.1m and peak heading errors of 2.5°. This trend of increasing errors as the road adhesion coefficient decreases highlights the necessity of integrating an EKF-based real-time road condition recognition for control compensation.

Figure 17 illustrates the control system’s coordinated response to steering commands and speed: 1) The steering angle response (top graph) shows significant peaks (about ±8°) at t=30s and t=90s, corresponding to sharp turn commands in the path; 2) The speed response (bottom graph) demonstrates intelligent coordination with steering demands—when the steering angle increases (e.g., at t=30s), the speed automatically decreases from 1.8 m/s to 1.3 m/s (a reduction of about 28%), effectively reducing the risk of skidding; during straight segments (e.g., at t=60s), the speed returns to the optimal operating speed of 1.8 m/s. This negatively correlated coordination mechanism is key to ensuring vehicle stability.

To comprehensively evaluate the system’s performance across spatiotemporal dimensions, Figure 18 presents a unique three-dimensional trajectory perspective: 1) Spatial dimension (x-y plane): The actual trajectory (red curve) and the reference path (blue dashed line) projections highly overlap, visually demonstrating path tracking accuracy; 2) Temporal dimension (x-axis): The continuous control process over the full 120-second operation cycle is fully displayed; 3) Performance dimension (z-axis): The tracking error height (z-value) clearly quantifies control deviations, with green points (dry road surface) generally showing z-values below 0.03 m, while brown points (muddy road surface) rise to 0.05-0.08 m; 4) Key phenomenon: At t = 60 s, when the road condition abruptly changes from dry (green points) to wet (blue points), the error height (z) exhibits a brief spike (around 0.07 m). However, thanks to the EKF’s rapid detection and the PID controller’s real-time compensation, the error quickly converges within 0.5 seconds (indicated by the purple arrow). This multi-perspective analysis validates the system’s adaptability and robustness in dynamic farmland environments.

Figure 19’s violin plots quantify the distribution characteristics of lateral errors under different road surface conditions: 1) Dry surface (left) shows the most concentrated error distribution (narrow waist), with 95% of errors falling within the 0.02-0.06 m range (box range); 2) Wet surface (middle) has a slightly wider distribution, with errors mainly between 0.04-0.08 m; 3) Muddy surface (right) exhibits the most dispersed distribution (wide belly), with error range expanding to 0.03-0.12 m, and extreme values at the ends of the whiskers reaching 0.15 m. The statistical results indicate that the system meets the agricultural operation standard of lateral error ≤ 0.1 m in the vast majority of conditions (95%). The control system demonstrates optimal stability on dry surfaces (with concentrated distribution), while exhibiting the widest error range on muddy roads. This phenomenon primarily stems from the nonlinear intensification of tire force saturation under low adhesion conditions, which makes vehicle response to steering inputs more unpredictable and increases control complexity. Furthermore, vehicle vibrations caused by road surface irregularities under low adhesion conditions significantly exacerbate tracking precision interference.

After completing the simulation validation, a detailed field test deployment plan was developed to further evaluate the performance of the proposed algorithm in real-world conditions. The test plan includes selecting an appropriate test site--a test field located in Shuangfeng County, Loudi City, Hunan Province--to ensure safety and controllability. The vehicle employed in the experiment was the Nongjiapan 4LZ-1.5A full-feed crawler combine harvester, with a total mass of 1200 kg (see Table 4). This mass parameter was essentially consistent with the equivalent scaled mass parameter (1250 kg) in the simulation model, exhibiting an error margin within 4.2%. This ensured comparability between simulation and experiment in terms of inertial characteristics. The vehicle will be equipped with necessary sensors such as the Intel RealSense D415 depth camera, an IMU (Inertial Measurement Unit) for real-time measurement of the vehicle’s three-axis acceleration, angular velocity, roll/pitch angles, and an RTK dual-antenna system for real-time differential data output (Fan et al., 2019), among others, to monitor trajectory, speed, and steering angle data in real time. Sensor configuration parameters are listed in Table 5. The picture of the whole machine is shown in Figure 20. The testing will be conducted in phases: initially verifying the system’s basic functions under simple conditions (such as straight lanes), then gradually progressing to more complex scenarios like curves and obstacles (referring here to field ridges) to assess the robustness and response speed of the PID controller. Data collection will cover different road adhesion coefficient (such as dry, wet, and muddy low-friction conditions) and will be compared and analyzed against simulation results. Additionally, the deployment plan incorporates safety protocols, including emergency braking mechanisms and a remote monitoring system, to ensure the safety of both the vehicle and personnel during testing. Actual operations will be carried out in rice paddies following typical harvesting paths (such as rectangular loops) with autonomous driving, evaluating the system’s overall performance and reliability in real crop environments and during extended operation periods.

As shown in Figure 21, the lateral error on dry, flat farmland varies over time. The error starts at 0.15 meters, then quickly decreases and stabilizes around 0.05 meters after the controller adjustment, remaining consistently below the 0.1-meter safety threshold. This verifies the stability of the PID control under simple conditions.

Figure 22 shows the changes in steering angle, speed, and lateral error during an upcoming turn at a field ridge. The steering angle reaches its peak (± 8°) at the curve, while the speed correspondingly decreases (from 1.8 m/s to 1.3 m/s), demonstrating the steering-speed coordination mechanism. The lateral error slightly increases at the curve (up to 0.1 m) but remains within a safe range. The control signals (steering angle and velocity) are filtered by a low-pass filter (cut-off frequency 2 Hz) to represent the low-frequency vehicle dynamics. The measured lateral error is raw data.

Figure 23 shows the lateral error and heading error on dry, wet, and muddy road surfaces. The errors increase as the road adhesion coefficient decreases, but the lateral error on muddy roads is still controlled around 0.1 meters (occasionally exceeding this but quickly corrected), and the heading error reaches a maximum of 2.5 degrees on muddy roads. When the road surface changes abruptly (at 40 seconds and 80 seconds), the controller completes adjustments within 0.5 seconds, and the errors quickly converge. As can be seen from Figure 23, the horizontal error has a trend of slow increase, which may be caused by the fixed deviation of GPS positioning signal or the cumulative effect of controller integration term. In the future work, this problem will be suppressed by online sensor calibration or the introduction of integration term limiter.

Figure 24 presents the lateral error distribution of each segment of the rectangular path using box plots. The errors in the straight segments (0.02-0.05 m) are significantly smaller than those in the curved segments (0.05-0.10 m). All errors are below 0.1 m, meeting the requirements for agricultural operations.

Figure 25 illustrates the actual obstacle-avoidance path of an autonomous harvester within a 60 m × 40 m field. The blue dashed line represents the hybrid A* planning path, while the orange solid line denotes the actual trajectory. The system successfully navigated around six obstacles, achieving an average lateral error of<0.15 m and a maximum error of<0.25 m. With a 100% obstacle avoidance success rate, this demonstrates the practical feasibility of the proposed framework. This figure provides visual evidence of the autonomous driving system’s actual obstacle avoidance capabilities as described in the paper.

These results indicate that the autonomous driving framework proposed in the paper basically meets the design objectives in real farmland environments, especially demonstrating excellent performance in path tracking accuracy and control response speed. However, its robustness under extreme road conditions still requires further optimization.

This study retains certain limitations that warrant further refinement in future work:

Insufficient consideration of crop dynamics: The simulations and experiments primarily focused on harvested or bare plots, failing to systematically account for the additional resistance imposed by standing crops on vehicle movement and its impact on dynamic characteristics and control stability. In practical operations, variations in crop density, height, and stalk strength may induce additional longitudinal/lateral load changes, thereby compromising path-following accuracy.

Limited applicability of model equivalence under extreme conditions: Although parameter adjustments achieved core dynamic equivalence between the wheeled model and tracked harvesters, the non-linear characteristics unique to tracks—such as slippage and sinking—may exhibit significant deviations in deep mud or high-resistance crop areas. The predictive capability of the existing model under such extreme conditions requires further validation.

Environmental perception and dynamic planning capabilities require enhancement: The current system primarily addresses static obstacles and fixed paths, lacking integrated real-time detection and avoidance of dynamic obstacles. It also fails to account for the dynamic impact of crop growth conditions on navigable areas.

Based on the findings of this study and the aforementioned limitations, subsequent research may proceed along the following lines:

Establish a crop-vehicle coupled dynamics model by incorporating a crop resistance module into simulations, thereby more accurately reflecting vehicle-environment interactions during harvesting operations.

Develop control strategies adaptable to crop conditions, either by enhancing PID controllers with real-time resistance estimation or introducing model predictive control (MPC), to improve tracking robustness during harvesting tasks.

Enhance dynamic environmental perception and planning capabilities by integrating multi-sensor information to achieve dynamic obstacle recognition and real-time passable area updates, subsequently investigating real-time replanning algorithms.

Conduct long-term field validation across multiple crops and operating conditions through large-scale field trials under diverse crop types and cultivation patterns, further verifying the system’s reliability and practicality within complex real-world agricultural environments.