2.1. Overview

The electric crawler-type weeding robot used in the experiment is shown in Fig. 1. The weed mower was a two-rotor type. The dimensions of the robot were 1650 mm (L) \(\times\) 935 mm (W) \(\times\) 900 mm (H), and its mass was approximately 150 kg. A LiFePO\(_{4}\) battery (DC36 V, 40 Ah) was used as the power source. The robot moved using left and right crawlers driven by independent brushless motors with a gearhead. The weed mower was driven using the same brushless motor without a gear head. These motors had a rated torque of 20 Nm and a rated output of 480 W.

Normally, a farmer can operate a robot in a radio-controlled manner. The robot used an RTK-GNSS (Vision-RTK 2, Fixposition Co., Ltd.) for autonomous operation. When the robot operated in autonomous mode, a notebook PC was used as the host controller. The development environment of the control framework was Robot Operation System (ROS) Noetic. The PC obtained the position and orientation from the RTK-GNSS through Ethernet, processed the data, and sent the control command through the controller area network (CAN).

The robot can mechanically change its tread between 700 mm and 900 mm. Furthermore, the robot can change the height of the left and right crawlers to adjust its attitude toward the sloped terrain.

2.2. Control Method

The speed control method for the robot involved the transmission of the target speed commands to the left and right motors via the CAN bus, thereby inducing rotation in the crawlers and allowing forward movement of the robot ⁶. The robot steering method used a two-wheeled vehicle model. The angular velocity \(\omega\) of the vehicle is calculated using Eq. \(\eqref{eq:1}\) as follows:

The speed of the right and left crawler is determined by angular velocity \(\omega\). If the value of \(\omega\) is greater than zero, the vehicle is steering to the left. While if the value of \(\omega\) is less than zero, the vehicle is steering to the right.

The angular velocity \(\omega\) is calculated based on the lateral deviation \(y\) and directional deviation \(\theta\) between the vehicle’s current position and target line. As shown in Fig. 2, the target position was located at the end of the straight target line. Distance \(L\) is the distance from the current position to target position. The angle \(\phi\) is the angle from the current position to target point. The target angular velocity \(\omega_{\textrm{d}}\) for the vehicle to turn toward the target position is expressed by Eq. \(\eqref{eq:4}\) as follows:

In this equation, \(\Delta T\) denotes the sampling time of the control process. When angle \(\phi\) is a small value, \(\sin^{-1}(y/L) \simeq {y}/{L}\) can be approximated. Based on the above equations, when given the target speed \(v_{\mathrm{d}}\), speeds of the left crawler \(v_{\mathrm{Ld}}\) and right crawler \(v_{\mathrm{Rd}}\) are calculated with the target angular velocity \(\omega_{\mathrm{d}}\) using Eqs. \(\eqref{eq:5}\) and \(\eqref{eq:6}\) as follows:

The target speeds for the left crawler \(v_{\mathrm{Ld}}\) and right crawler \(v_{\mathrm{Rd}}\) can be calculated, and the referred command value can then be transmitted via the CAN.

2.3. Crawler Height Adjustment

As shown in Fig. 3, the heights of the left and right crawlers could be adjusted by changing the fixed angle of the chain case. This chain case was positioned on both the outer sides of the robot’s body and secured in place with screws on each side, allowing for height adjustment. The robot maintained a balanced load distribution by simultaneously changing its rolling angle and chassis width. In the experiment, two crawler heights were selected for autonomous operation on the slopes.

2.4. Route Planning

As shown in Fig. 4, the experimental area was defined by a rectangle. The area was divided into multiple columns for backward and forward operations based on the width of the mower. The rectangular area designated for the experiment was 16 m long and 2.4 m wide. Given the effective width (0.7 m), the area was divided into three lanes. Each lane was assigned start and end points. After reaching the endpoint, the vehicle reverted to the starting point of the adjacent lane and began operation in that new lane.

3.1. Autonomous Navigation Experiment

The field experiment was conducted in a weed area on a steep slope, located in Yabu-cho, Yabu-shi, Hyogo Prefecture, on August 28, 2025 (Fig. 5).

As shown in Fig. 5, the slope of the weed area was 35°. The slope of the experimental area varied only slightly; therefore, the resulting effect was minor. The crawler robot started from the starting point, following the routes described in Section 2.4. Two autonomous navigation experiments were conducted using robots with different roll angles. Owing to mechanical constraints, only roll angles of 10.5° and 17.7° were selected as feasible experimental conditions to represent the optimal outcome achievable within this setup. Other angles were not included in this experiment owing to safety considerations. Experiment 1 was conducted using a robot with a roll angle of 10.5°. Experiment 2 was conducted with a roll angle of 17.7°. In both the experiments, the mowing load was set to the same level. However, this setting was made nonfunctional for safety reasons.

3.2. Evaluation

To evaluate the autonomous travel of the crawler robot, the path-following error and area coverage were evaluated.

3.2.2. Area Coverage

To further evaluate the performance of the weeding operation of the crawler robot, the coverage area of each experiment was evaluated.

The area coverage was determined based on the actual trajectories of the robot and width \(w\). Let the actual GNSS path be represented as a sequence of positions \(P = \sum p_{i}\). Between every pair of adjacent positions, the rectangular area \(R_{i}\) representing weeding coverage was calculated from the distance between two adjacent positions and the mower’s width using the following equation:

\begin{equation} R_{i} = \sqrt{\bigl(x_{i+1} - x_{i}\bigr)^{2} + \bigl(y_{i+1} - y_{i}\bigr)^{2}} \times w. \label{eq:9} \end{equation}

Let the area after merging the first \(i\) polygons be denoted as \(A_{\cup}^{(i)}\). When adding the next polygon with area \(R_{i+1}\), the updated union area \(A_{\cup}^{(i+1)}\) is expressed as follows:

\begin{align} A_{\cup}^{(i+1)} &= A_{\cup}^{(i)} + R_{i+1} - A_{\mathrm{overlap}}^{(i)}, \label{eq:10} \\ \end{align}

\begin{align} A_{\mathrm{overlap}}^{(i)} &= \mathit{Area}\left(R_{i+1} \cap \bigcup_{j=1}^{i} R_{j}\right), \label{eq:11} \end{align}

where \(A_{\cup}^{(0)} = 0\) (initial empty set), and \(A_{\mathrm{overlap}}^{(i)}\) denotes the area of overlap between the new rectangle area \(R_i\) and the existing union. The overlapped area coverage is the new area obtained by subtracting any overlapping regions with previously merged polygons, which is calculated using the plane sweeping algorithm in the Shapely Python library.

4.1. Path Following Error

The autonomous driving trajectories in Experiments 1 and 2 using the robot are shown in Fig. 6. The travel trajectory was projected onto a horizontal plane and did not exhibit any changes in elevation. As shown in Fig. 6, the target path is plotted as a dotted line, and the actual travel path recorded by the RTK-GNSS is plotted as a solid line. Both the autonomous driving paths were closely aligned with the target paths. It was demonstrated that the crawler-type electric weeding robot could follow the target path and finish the operation well on the slope area.

Table 1 lists the deviation errors of the two experiments. The average deviations in Experiments 1 and 2 were 0.04 m and 0.02 m, respectively. After the height adjustment of the roll angle for the robot, the deviation error decreased for better performance. The maximum deviation errors were 0.15 m and 0.13 m in two experiments. These errors frequently occur in the reverse path owing to slippage. Experimental observations showed that a roll angle of 17.7° resulted in slightly less slippage than a roll angle of 10.5° during backward motion on a slope. Because of the lack of active control in the roll direction, the observed difference between the tested conditions was marginal.

4.2. Area Coverage Results

Figure 7 shows the area coverage of the weeding performance during autonomous driving following the target path. The black lines represent the actual travel paths of the mowing-robot-recorded GNSS positioning system. Green lines indicate the length of the mower’s width and recorded position.

The overall coverage areas are listed in Table 2. The target coverage areas in both experiments were 34.24 m\(^2\). The actual coverage areas in Experiments 1 and 2 were 33.96 m\(^2\) and 34.0 m\(^2\), respectively. Both coverage percentages were greater than 99%. The weeding performance of the electric vehicle was acceptable in the sloped area.