The agricultural industry is rapidly embracing technological advancements to increase its efficiency and productivity. One of the most promising technologies involves the application of autonomous mobile robots in smart greenhouses. These robots can perform various tasks such as monitoring plant growth, harvesting crops, and spraying pesticides and fertilizers. They can also reduce labor costs, increase crop yields, and improve the overall quality of crops. Smart greenhouses equipped with sensors and automation systems, including temperature, humidity, and lighting systems, are used to control the environment. The integration of autonomous mobile robots can further enhance the efficiency and precision of greenhouse operations. These robots can navigate greenhouses using various sensors and algorithms and make decisions based on data collected from the environment and plants [1].

The task of an autonomous mobile robot in a smart greenhouse is to assist with various operations, such as monitoring plant growth, harvesting crops, spraying pesticides and fertilizers, and performing other tasks related to crop management. Smart greenhouses are dynamic environments that require continuous monitoring and adjustment to optimize plant growth and manage resources [2–4].

Several task-specific robots are used for monitoring, spraying, and harvesting. A task-specific robot may be limited to performing only one specific task or a set of tasks and may not be able to adapt to changes in the greenhouse environment or new requirements. For example, a robot designed specifically for harvesting may not be able to perform other tasks such as spraying pesticides or monitoring plant growth. This can limit the efficiency and versatility of the robot and may lead to the requirement of additional robots for different tasks.

In contrast, a robot platform has several advantages over a task-specific robot in smart greenhouse applications. Robot platforms can be customized and reconfigured for different applications and tasks, making them more versatile than task-specific robots. This reduces the need for specialized robots for each task, which can be costly and inefficient.

The increasing demand for efficient agricultural practices has led to the emergence of smart farms, where automated systems play a vital role in optimizing crop production. However, the successful deployment of robots in such environments remains challenging because of the narrow and confined spaces they often encounter. To address this issue, we propose a mobile-robot platform that is specifically engineered to maneuver seamlessly in these demanding settings. Inspired by innovative techniques employed in Korean smart farms, our platform aims to enhance navigation capabilities and improve overall productivity.

To achieve this goal, we conducted a comprehensive review of various types of mobile robots. As shown in Figure 1, we designed the mobile robot concept, mechanical system, electrical system, robot kinematics, dynamics, and path-tracking controller and evaluated the robot’s performance. The mechanical strength of the robot platform was verified using the finite element method (FEM). A path-tracking control was installed on the robot to evaluate its performance. Several path-tracking control methods using proportional-integral-derivative (PID) control for autonomous robots have been discussed. To overcome the limitations of PID control, an adaptive fuzzy PID control method was proposed for path tracking. An experimental setup was built using a robot platform developed for path-tracking control. The PID controller and proposed fuzzy PD controller were tested for path tracking of the robot platform. We compared the performances of both the PD controllers with different parameters and the proposed fuzzy PD controller applied to the robot platform. The effectiveness of the developed robot was experimentally assessed.

2.1 Types of Mobile Robot Platforms

Navigating inside a smart greenhouse can be challenging for a mobile robot owing to the narrow and constrained environment. The robot must be able to move between rows of crops, navigate obstacles, and avoid damage to plants and other objects in a greenhouse [5].

The space available for a mobile robot in a smart greenhouse is often limited, and the paths between the rows of crops are narrow. This means that the robot must be designed to have a small footprint and be able to maneuver in tight spaces without damaging plants or other objects. The robot must also be able to avoid obstacles, such as plant stems, soil mounds, and irrigation hoses.

We present a review of several types of robot platforms to identify the most suitable type for smart greenhouses. Mobile robots are classified into five typical models based on their steering system design—four-wheel-differential drive, two-wheel-differential drive, car-like, steerable drive, and two-wheel-steerable drive—considering factors such as the purpose of operation, complexity, manufacturing cost, and degree of maneuverability [6–10]. This classification is illustrated in Figure 2.

As shown in Figure 2(a), the most common model among the five types of mobile-robot platforms is the steerable drive, which utilizes only one steerable-drive wheel to change the pose of the robot in terms of direction and position. The key advantage of this model is its precise tracking capability. However, the robot may face difficulty in changing direction in a nonmoving state because it has only one active wheel. To overcome this limitation, a two-wheel-steerable drive offering high rotational flexibility can be used, as shown in Figure 2(b). However, the cost of this model should be carefully considered because the steerable drive motor is relatively expensive, particularly for agricultural applications.

Another widely used model is the car-like model shown in Figure 2(c). This model uses a motor to drive the two rear wheels and propel the vehicle forward or backward. A steering mechanism based on Ackerman geometry is controlled by a steering motor, which is used to change the direction of the robot. This model has a huge advantage in stabilizing the linear velocity; however, as a car, it cannot drive sideways. Although this type is widely used in agriculture, it is not suitable for smart greenhouses, which have narrow and constrained environments.

A useful robot model for moving in a tight space is the differential-steering model, which uses independent drive wheels for movement. As each wheel can be driven at a different speed, the direction of the robot can be controlled.

With a similar operating design, the four-wheel-differential drive model shown in Figure 2(e) is easier to build in terms of the motor and driver selection [11]. However, the four-wheel differential drive model has difficulty synchronizing the speeds of the left and right wheels. Because of the same reference speed for the left or right pairs of wheels, one of them will skid if the other rotates faster. When a robot needs to move straight, the speed asynchrony of all the wheels makes it more difficult for the vehicle to maintain its direction in a smart greenhouse.

A mobile robot platform can be customized and reconfigured for various applications and tasks, making it a versatile tool for greenhouse applications. Therefore, we propose a mobile robot platform based on a two-wheel-differential drive that offers high maneuverability and simplicity of design.

2.2 Design of Mobile Robot Platform Structure

3.1 Path-Tracking Control for Autonomous Robot

The development of a path-tracking control system is of significant importance in the realm of autonomous mobile robots driven by wheels. Traditionally, a smooth path is planned for conventional mobile robots by incorporating curvature continuity, whereas a separate path-tracking controller is designed to mitigate the occurrence of path errors during navigation. However, executing smooth path planning in real time poses challenges owing to the computational burden. Furthermore, conventional path-tracking algorithms often lead to unpredictable tracking motion when faced with substantial path errors.

Regardless of the type of sensor used, this algorithm relies on the feedback signal from a position sensor installed on the body of the robot to guide it closer to the path [12–14].

Previous studies have demonstrated the effectiveness of conventional PID control techniques for path-tracking robots. Fuzzy logic and sliding-mode controls have also been proposed.

Fuzzy logic control has been successfully designed for path-tracking control [15]. Another approach is to use a sliding-mode control [16]. Neural networks, reinforcement learning, and vector quantization have been studied for a path-tracking robot running on a path with tight curves at high speeds.

Balaji et al. [17] presented the roles of the proportional, integral, and derivative terms of a PID controller in path-tracking control. For example, with a large proportional gain value, the robot can respond quickly when the path changes suddenly but oscillates widely in straight lines. The derivative term of a PID controller is known for its overshoot reduction function and transient response improvement caused by sudden changes in the path trajectory. In the presence of an integral term in a PID controller, the performance of the mobile robot can become unstable because of the instability of the reference value for the entire system. Therefore, the PD controller for path tracking was implemented with a proportional gain that could be adjusted, whereas the derivative gain should be selected appropriately and should remain unchanged.

3.2 Two PID Controllers for Path-Tracking

First, we developed a path-tracking control algorithm for the proposed autonomous robot platform using a PID controller, which is the most commonly used controller in automatic systems.

The motor driver of the mobile robot controls the rotating velocity of each wheel, which depends on the left and right velocity commands u_L and u_R, respectively. According to the robot’s kinematics, the left and right velocities can be controlled by the robot’s velocity and angular velocity, u_v and u_e, respectively, which can be written as

Therefore, the path-tracking control algorithm consists of two controllers: a linear velocity controller and an angular velocity controller. The PID controller of the angular velocity can be replaced by a PD controller to reduce wounding by the integral term.

Thus, the equation of linear velocity and angular velocity control can be written as

where [u_v u_e] is the velocity control command and the angular velocity control command; [K_pɛ K_dɛ] are the parameters of the PD controllers of the path-tracking position; and ɛ and e_ɛ = −ɛ are the path position and position error, respectively.

Although the PID parameters can be tuned for a specific robot condition in a given environment, such conditions may change over time. In such cases, fixed PID controllers may not be able to adapt to the new conditions. Therefore, an intelligent control method that can adapt to the changing conditions is required.

3.3 Adaptive Fuzzy PID Control for Path-Tracking

Fuzzy logic is a form of many-valued logic in which the true values of variables are employed. It is widely used in various applications [18, 19].

The fuzzy logic controller (FLC) is widely known for its ability to supervise nonlinear systems and adjust the PID parameters to optimize the state of the system [20–23].

In this study, we propose an adaptive fuzzy PID controller that includes a PID controller for linear velocity and a PD controller with an FLC for angular velocity. The FLC adjusts the proportional gain of the PD controller to accommodate the changes in the system. To achieve this adaptation, we utilized the Sugeno fuzzy controller, which is efficient for computational processing, as shown in Figure 6.

The derivative term K_dɛ in the PD controller plays a crucial role in facilitating effective direction changes for the robot. However, if the gain K_dɛ is large, the system becomes highly sensitive to noise. To enhance the response of the robot to changes in the path curve, a fuzzy logic controller is implemented to adjust the gain K_pɛ, whereas the derivative gain K_dɛ is chosen appropriately and kept constant.

where k_pɛ ∈ [0, 1] is the output signal of the fuzzy controller.

The fuzzy rule for the controller is built using nine rules, in which DE = {NDE, ZDE, PDE}, E = {NE, ZE, PE}, and KP = {Z, M, B} are the sets of linguistic variables of the fuzzy inputs and outputs (e_ɛ, de_ɛ, and k_pɛ), respectively. Each rule is created as a tracking rule.

where n = 9 is the number of fuzzy control rules and i = j = 1, 2, 3.

4.1 Experimental Setup

The developed autonomous mobile robot platform is depicted in Figure 7, with the robot body having a width of 0.56 m and a length of 0.84 m. The diameter of each wheel was 0.3 m. The robot weighed 125 kg without a payload. To enable path tracking, a magnetic position sensor with a measurement range of 14 cm and a measurement resolution of 0.01 cm was installed in front of the robot. The velocities of the left and right wheels were obtained from digital signals generated by the Hall-effect sensors embedded inside two BLDC motors, with the maximum corresponding speed of the wheels limited to 60 rpm. Open-source Kalman filters were applied to improve the velocity measurement accuracy and eliminate unwanted velocity noise, with measurement uncertainty, estimation uncertainty, and process variance set to 2, 2, and 0.01, respectively.

4.2 Path-Ttracking Control Test

In this study, the performance of the developed mobile robot with path-tracking controllers was evaluated by tracking a predefined path, as illustrated in Figure 8. The robot operated at a constant speed of 0.157 m/s, corresponding to a wheel speed of 10 rpm. The path consisted of a rectangular shape with four corners, namely C1, C2, C3, and C4, as shown in Figure 8. The experimental evaluation of the robot involved traversing the path twice, corresponding to the eight corners.

In this paper, we present the results of three experiments aimed at tuning the proportional and derivative gains of a PD controller for angular velocity. The first experiment utilized gains of 0.2 and 0.25, whereas the second experiment used gains of 0.6 and 0.25, and the third experiment used gains of 1.0 and 0.25. The purpose of these experiments was to investigate the effect of varying the proportional gain on controller performance while keeping the derivative gain constant. Experiments were conducted using a robot test setup, and the results were analyzed to determine the optimal parameters of the PD controller for angular velocity control.

This study also investigated the performance of the proposed fuzzy PD controller for the angular velocity. To determine the parameters K_pɛMin, K_pɛMax, and K_dɛ in Eq. (3), experiments were conducted in which the robot ran on a closed path at a reference speed of 0.157 m/s. Specifically, we followed the tracking steps outlined below:

• Step 1: Gradually increase K_pɛ until the robot can follow the straight path. However, when the direction of the path changed by 45° from the original direction, the robot was unable to follow the path. The value of K_pɛMin at this point was considered K pɛMin.

• Step 2: Increase K_dɛ until the robot could follow the entire closed path. K_dɛ can be chosen to be larger, but only to the extent that it allows the robot to change direction more quickly.

• Step 3: Maintain K_dɛ constant and increase K_pɛ until the robot oscillates strongly on a straight path. The value of K pɛ at this point is considered K_pɛMax, ensuring that the robot can follow a closed path.

To evaluate the performance of the robot platform, we tested it using a PD controller with different parameter values, as well as the proposed fuzzy PD controller. Specifically, we ran the robot in two rounds along a predefined path at a speed of 0.157 m/s, which corresponded to a wheel speed of 10 rpm. The experimental results are presented in Figure 9, where we measured the lateral distance error of the robot from the center of the path.

The results indicate that when K_pɛ is the smallest (0.2), the robot exhibits the worst ability to follow the path at the corners, as expressed by the sky blue solid line in Figure 9, which shows a high overshoot at the corners. Conversely, when K_pɛ is the largest (1.0), the robot exhibits the best ability to follow the path at the corners, but it oscillates much more widely, even on a straight path, as indicated by the solid red line in Figure 9. The optimal value of K_pɛ was found to be 0.6, as indicated by the green solid line in Figure 9, which exhibits a medium overshoot at the corner and a small oscillation along the straight path, although the overshoot is still larger than that of the smallest K_pɛ(0.2).

Figure 9 also shows a zoomed-in view of the lateral distance error in the straight path (Box A), where the proposed fuzzy PD control outperforms the PD control. Specifically, the overshoot at the corner by the fuzzy PD control was as small as that of the smallest K_pɛ(0.2), and the oscillation along the straight path was smaller than that of the PD control, as indicated by the blue solid line in Figure 9.

From the results presented in Figure 9, we measured the performance indices, as listed in Table 1. Based on the results of the root mean squared errors (RMSEs), we found that the larger the K_pɛ, the smaller the error at which the robot tracks the path. However, it has the disadvantage of making the system oscillate strongly on the straight path, where the maximal peak-to-peak error ripple is the greatest. With the smallest K_pɛ, the system becomes the most stable, and its ability to track the path worsens at the corners.

In conclusion, we proposed a novel mobile robot platform engineered to excel in navigating narrow and cramped spaces, such as those found in smart farms. The proposed mobile-robot platform demonstrated remarkable adaptability and agility through the incorporation of an innovative robot design and advanced adaptive control. We implemented path-tracking controls and proposed an adaptive fuzzy PID controller to evaluate the performance of the robot platform. We tested and compared the performance of the PD controller with different parameters to that of the proposed fuzzy PD controller on a robot platform. Promising results obtained from preliminary experiments indicate that this platform has the potential to customize the field of agricultural robotics. Further research and development are required to optimize the performance of the platform and evaluate its real-world applicability. Nevertheless, we remain confident that our proposal represents a significant step forward in enhancing the efficiency and productivity of smart farming practices.

Table 1. Experimental results by performance indexes (unit: cm).