Development of a multipurpose trolley for greenhouse cultivation

Abstract

Japan’s agricultural sector faces serious challenges due to a declining and aging workforce. Greenhouse cultivation requires intensive labor in confined spaces, making efficiency improvement and reduced physical burden urgent. This study proposes a solution that converts a commercially available harvesting trolley into an electric multifunctional platform capable of both hands free assistance and autonomous operation. Key technological contributions include a travel compensation control system that enhances operability and positional accuracy and a torque transmission mechanism that adapts to multiple tasks. The evaluation experiments showed reduced positional errors during movement in narrow greenhouse aisles. In addition, the practical utility of the two modes, human following and autonomous movement, was confirmed, indicating potential for agricultural support.

1. Introduction

Japan’s agricultural industry is facing a severe labor shortage due to a declining and aging workforce (Ministry of Agriculture, Forestry, and Fisheries [MAFF], 2024b). Greenhouse cultivation, with its confined workspace and reliance on manual labor, presents a significant challenge (e-Stat, 2022). The work is often physically demanding and potentially hazardous, especially in the hot and humid environment within greenhouses (MAFF, 2023). To address this, the MAFF has promoted the use of IoT and robots to improve efficiency and reduce labor (MAFF, 2024a). In addition to promoting the automation of tractors (Nagasaki, 2021), robots that support greenhouse cultivation are also being developed. Existing robots often rely on rails or guides, limiting their adaptability to existing greenhouses (Hayashi et al., 2012; Imou, 2017). This study focuses on developing a small guide-free support trolley that can adapt to the greenhouse environment.

Our initial objective is to support tasks involving a trolley, such as leaf cutting and harvesting. These tasks are conventionally performed while manually operating the trolley, which can be unstable on uneven greenhouse terrain, thereby hindering worker concentration and efficiency. To achieve hands-free operation, we adopted methods from previous studies on human-following vehicles (Sasaki et al., 2011; Yorozu, 2021) that utilized worker and furrow detection for stable movement.

Another significant demand is for unmanned pesticide spraying. Carrying heavy spraying equipment poses a considerable burden and exposes workers to potentially harmful chemicals (Kim et al., 2015). Autonomous agricultural support vehicles and robots are being actively researched (Fukao, 2021), and we aim to automate pesticide spraying using a sprayer-equipped support trolley. Given the limited area and numerous obstacles within a greenhouse, autonomous movement using GPS is impractical (Nagasaki, 2021). Visual SLAM, while a potential method, is susceptible to external factors (Bai et al., 2023). Therefore, we employ a range sensor-based map generation approach that enables precise destination setting and obstacle recognition (Fujinaga et al., 2021; Mousazadeh, 2013; Xiong et al., 2019). Other relevant studies include autonomous movements based on ridge recognition (Åstrand, 2005) and obstacle avoidance using a combination of range sensors and cameras (Han et al., 2023).

This study proposes a multipurpose trolley for greenhouse cultivation, adaptable for various tasks ranging from harvesting to pesticide spraying. By utilizing a commercially available trolley as the base, we aim to reduce costs and improve adaptability. The trolley supports manned tasks through hands-free operations, allowing workers to focus on their tasks while the trolley follows them. Autonomous movement using a range sensor enables transportation and pesticide spraying, further reducing labor and ensuring worker safety. The trolley maintains its posture and position between benches, enhancing its stability and operability on uneven terrain. In addition, an automated charging system at a charging station streamlines battery management. We have implemented and verified the proposed functions through experiments, the results of which are presented in this paper.

2. Materials and methods

The developed agricultural work support trolley aims to reduce the workload in greenhouse environments. It offers direct support to workers through remote control and hands-free human-following modes. Additionally, it automates tasks using an autonomous-movement mode and automates charging in an entering operation mode. Travel compensation control ensures stable movement across various modes on uneven terrain. This section describes the specifications of the multipurpose trolley, operating system, and control modes.

2.1. Configuration of the prototype trolley

The prototype trolley and the system configuration are shown in Figs. 1 and 2, respectively. The prototype trolley is based on the commercial trolley (Aisai-go SW-312, HARAX Coop., Japan) and is equipped with a control unit (PC and microcomputer), a laser range finder (LRF) as a range sensor for person-following and autonomous driving, an IMU for accurate odometry, a magnetic resonance-type contactless charger for automated charging, and a geared motor with an encoder, drive wheel, and transmission system for adjusting the driving torque. It uses a 25.1V lithium-ion battery and can run for 90 min in second gear with no load. In addition, the side with the LRF is the front, which is the side where the worker walks. The control system utilizes ROS 1, operating on a node-by-node basis and enabling distributed data exchange between independent systems (Open Robotics, 2016). All the modes operate in parallel, switching based on the output signal.

Fig. 1 An overview of the prototype trolley

Fig. 2 The system configuration

2.2. Transmission

To accommodate the diverse tasks within a greenhouse, each requiring a different optimal torque, the trolley incorporates a three-stage transmission system. The first gear provides high torque for heavy loads, the third gear allows for speeds matching the human walking pace (4 km/h), and the second gear offers a performance between the two. All the gears are module 2 spur gears, with the first gear having 12 and 30 teeth, the second gear having 18 and 24 teeth, and the third gear having 24 and 18 teeth. The transmission specifications are listed in Table 1. Gear-shifting is achieved by a servomotor moving the shaft laterally to mesh the desired gear combination, aided by a linear bush for lateral movement and a sprocket for maintaining power transmission during shifting. The transmission, when set in the second gear, is illustrated in Fig. 3.

Table 1 Transmission specifications

Gear	Reduction ratio	Maximum load (kg)	Max speed (m/s)
1st	240	144	0.28
2nd	130	75	0.53
3rd	73.1	42	0.94

Fig. 3 The transmission structure

2.3. Control systems

2.3.1. Object detection

The LRF detects feet, poles, walls, and charging stations used for the person-following mode, travel compensation control, and entry operation mode. The detection uses a node of the “Obstacledetector_packege”, one of the ROS packages (Przybyla, 2017). This package identifies obstacles as circles or lines from 2D point cloud data. A custom node of our design differentiates these obstacles into feet, poles, walls, and charging stations, based on threshold values. The feet and poles are distinguished by the diameter of the circle, with objects of 0.02–0.04 m being considered as the pole and those of 0.08–0.15m being considered as the feet. The charging station is marked by a board 0.55m wide as a landmark. Therefore, obstacles in a line of 0.4–0.7 m are judged to be charging stations. Linear obstacles larger than that are considered walls.

2.3.2. Human-following mode

In the human-following mode, the trolley follows a person based on feet detection using the LRF. It maintains a set distance from the detected feet using a proportional-derivative control approach, adjusting its speed and turning based on the error from the set distance and travel compensation control signals. The procedure first detects the feet and then measures the distance from the trolley. Next, the error with the set distance is calculated, and finally, the control signal is output according to the error. The control signals are shown in Eqs. (1) and (2) as follows. For this function, we assume that there is only one worker and that the system targets the nearest foot.

  

$$v_L\left(t\right)=\{K_p{y}_e\left(t\right)+K_d\frac{d}{dt}{y}_e\left(t\right)\}\{1+e_c\left(t\right)\}$$(1)

  

$$v_R\left(t\right)=\{K_p{y}_e\left(t\right)+K_d\frac{d}{dt}{y}_e\left(t\right)\}\{1+e_c\left(t\right)\}$$(2)

Where v_L (V) is the control signal for the left wheel, v_R (V) is the control signal for the right wheel, K_p is the proportional gain, K_d is the differential gain, y_e is the error from the set distance, $\frac{d}{dt}$ y_e is the time derivative of y_e, and e_c is the control signal from the travel compensation control. The error from the target distance is reflected in the translational speed, and the turning component is handled by the travel compensation control described below. Human following is achieved by adjusting the gains K_p and K_d in Eqs. (1) and (2). It can handle not only a simple walk but also a walk in which the person stops each time.

2.3.3. Autonomous-movement mode

Autonomous movement within a greenhouse involves self-localization based on environmental recognition and route generation to a designated destination. Although ROS provides the Navigation Stack for autonomous movement (Open Robotics, 2016), modifications are necessary for the greenhouse environment. This section describes the processes and algorithms used to achieve autonomous movement, including these improvements.

Self-localization based on environmental recognition requires an environmental map, which is generated using gmapping (Open Robotics, 2019), a package in the slam_gmapping package group for generating environmental maps. An environmental map showing obstacles is generated by manually driving trolleys and acquiring sensor data. This is a method known as SLAM, which estimates the trolley’s self-position and generates an environmental map by piecing together the surrounding environment as the trolley moves. The map was created by moving a trolley through the driving environment. As an example, a test field simulating the environment in a greenhouse cultivation and the generated environmental map are shown in Figs. 4 and 5, respectively. A part of the test field has lanes that simulate cultivation benches covered with skirts. For self-localization, the emcl2 package, incorporating a method by Ueda et al. (2004), was employed. This method acquires sensor data sequentially even when the trolley is stationary, thereby enabling accurate self-localization to be continued.

Fig. 4 A test field

Fig. 5 The generated environment map

For autonomous movement, an environmental map is used to set destinations and plan routes. After the destination is set, the system generates a route from the self-position1 to the destination. This comprises two parts: a global route plan from the environmental map and a local route plan from the surrounding dynamic obstacles. The system generates a cost map that considers the risk of collision with obstacles on the map, with higher costs near obstacles and lower costs at points farther away. By adopting a route at a low cost, a safe route that avoids collision between the trolley and obstacles is generated. When a new obstacle is detected nearby, the cost function is updated accordingly, and a path is dynamically replanned to avoid the obstacle. The destination can be set manually on Rviz, an ROS visualization tool, or by sending the goal data directly. In this case, multiple destinations can be sent in sequence, setting them as waypoints and allowing the trolley to travel along any route. In this process, the cost is adjusted to generate a path that reaches the destination while maintaining a distance that avoids collisions with obstacles.

Furthermore, by mounting and operating equipment on the trolley while traveling along any route, agricultural work, such as spraying pesticides, can be automated.

2.3.4. Entering operation mode

This mode automates the final alignment with the charging station. After autonomous movement to the charging station zone, the trolley aligns its wireless charging unit with the station using a magnetic field resonance method (Kurs, 2007). An overview of the charging station is shown in Fig. 6. The wireless charger can charge the battery from 22.5V (empty) to 25.1V (fully charged) in 2.5 h. The charging station is marked with a board, and the trolley estimates its position relative to this mark for accurate alignment. The entering operation mode enables an accurate approach because it continuously calculates and corrects its positional error relative to the charging station. The alignment is performed in the following steps.

1) Draw the center line B that intersects at a right angle with the midpoint of the line A connecting both ends of the plate, which is the mark of the charging station.

2) Draw the line C that passes through the trolley and intersects at a right angle with the center line B.

3) Calculate the distance from the trolley position to the intersection of the lines B and C, which is the error x_e, the distance from the trolley position to the charging station is the error y_e, and the error in the trolley’s approach angle relative to the line B is θ_e.

4) Output a control signal calculated using each error as follows.

  

$$v_L\left(t\right)=K_{py}{y}_e\left(t\right)+K_{px}{x}_e\left(t\right)-K_{\theta }{\theta }_e\left(t\right)$$(3)

  

$$v_R\left(t\right)=K_{py}{y}_e\left(t\right)-K_{px}{x}_e\left(t\right)+K_{\theta }{\theta }_e\left(t\right)$$(4)

Where v_L (V) is the control signal for the left wheel, v_R (V) is the control signal for the right wheel, K_px, K_py, and K_θ are the proportional gain for each errors. The gains are adjusted by trial and error to achieve a smooth entry along the line B. The adjusted gains were set to K_px = 0.35, K_py = 0.2, K_θ = 0.7, respectively.

Fig. 6 An overview of the charging station

2.3.5. Travel compensation control

To ensure stable movement on uneven terrain, travel compensation control from the trolley’s position within the aisle between the benches and improves the straightness. This control is applied in the remote, human-following, and autonomous-movement modes. The system detects the poles of the bench/ the walls using the LRF, calculates the error from the specified path within the aisle between the benches, and derives a compensation control signal using fuzzy inference. The algorithm for calculating this control signal is as follows.

First, the poles and walls are detected by the LRF, and the error e_f(t) between the self-position of the trolley and the specified path is derived using the following equation.

  

$$e_f\left(t\right)=\frac{x_L\left(t\right)+x_R\left(t\right)}{2}-e_0$$(5)

Where x_R(t) is the distance to the pole/wall on the right side of the trolley, x_L(t)is the distance to the pole/wall on the left side, and e₀ is the distance from the center to the specified path.

Next, e_f(t) and its time derivative $\frac{d}{dt}$ e_f(t) are input into the triangular membership function with five labels, as shown in Fig. 7, and their respective grades are calculated. The fit grade μ_ij in the selected rule is calculated using Eq. (6).

  

$${\mu }_{ij}={\mu }_i\times {\mu }_j$$(6)

Fig. 7 Membership functions

Finally, the compensation signal e_c(t) is derived using a simplified inference method (Mizumoto, 1995) that takes the weighted average of the grades in Eq. (7) and the singleton variables w_ij in the rule table, as listed in Table 2.

  

$$e_c\left(t\right)=\frac{\sum _{}^{}{\mu }_{ij}{w}_{ij}}{\sum _{}^{}{\mu }_{ij}}$$(7)

Table 2 The rule table

		$\frac{d}{dt}$ ef(t)
		NB	NS	ZO	PS	PB
ef(t)	NB	0.25	0.25	0.19	0.125	0
	NS	0.25	0.19	0.125	0	−0.125
	ZO	0.19	0.125	0	−0.125	−0.19
	PS	0.125	0	−0.125	−0.19	−0.25
	PB	0	−0.125	−0.19	−0.25	−0.25

The control signals of the drive wheels, including the compensation control signal, are expressed by Eqs. (8) and (9). e_b(t) is the translation speed of the other control modes.

  

$$\begin{array}{c}\hfill v_L\left(t\right)=e_b\left(t\right){(1+e}_c\left(t\right)\hfill \end{array})$$(8)

  

$$\begin{array}{c}\hfill v_R\left(t\right)=e_b\left(t\right){(1-e}_c\left(t\right)\hfill \end{array})$$(9)

2.4. Evaluation methods

Several experiments were conducted to evaluate the system’s performance.

2.4.1. Evaluation of the travel compensation control

To confirm the stability effect of the travel compensation control, we conducted an experiment in which the trolley traveled between cultivations. The trolley traveled on lanes in an outdoor test field that simulated the interior of a greenhouse. The trolley was also compared in the remote-controlled and autonomous-movement modes. It traveled three times in each of the following modes: (A) remote-control mode, (B) autonomous-movement mode, and (C) autonomous-movement mode + travel compensation control.

The starting point is the center of the aisle entrance, and the path is 7 m long, with the center of the aisle between the benches as the set path. The set path is shown in Fig. 8. In the autonomous-movement mode, the trolley manually sets its position at the starting point and then manually sets the goal point to generate a route. In all the modes, the gearshift was in second gear. We recorded the trolley’s odometry and the pole positions from the LRF and calculated the driving trajectories to compare the effects of the travel compensation control.

Fig. 8 Experimental path of the travel compensation control

In addition, multiple routes were set and traveled to confirm the effect of the traveling route specification function in the corridor between the benches using the travel compensation control. The experimental environment was the same as that described above. The following paths were traveled three times between specific lanes; the travel path was set at the center and ±0.15 m from the center, as shown in Fig. 9. In this case, the operation mode is the autonomous-movement mode + travel compensation control, and the gearshift is in second gear. We recorded the trolley’s odometry and the pole positions from the LRF and calculated the driving trajectories to compare the effects of the travel compensation control.

Fig. 9 Experimental path of the route specification

2.4.2. Evaluation of the human-following mode

An experiment was conducted using a simulated continuous agricultural task to evaluate the worker-following mode. In the experiment, a trolley traveled in an outdoor test field and followed a human performing harvesting tasks. The harvesting tasks were simulated by retrieving 58 simulated crops placed in 29 planters on the cultivation benches, each containing 2 samples. This task was performed by traveling around the test field twice along the path between the cultivation benches, as shown in Fig. 10. The trolley traveled in the human-following mode, with travel compensation control applied to the straight sections, and the transmission was set to second gear.

Fig. 10 Experimental path of the human following mode

2.4.3. Evaluation of the autonomous-movement mode

The following experiment was conducted to verify the automation of multiple agricultural operations over a long period through battery charging. Unmanned agricultural tasks such as pesticide spraying and recharging were verified using the trolley’s autonomous-movement mode. The pesticide sprayer is mounted on a trolley and continuously sprays water over a designated area. The driving route and spraying sections are shown in Fig. 11. On the first lap, the trolley travels outside the cultivation bench and returns to the charging station. Once fully charged at the charging station, it leaves the charging station after a certain period. It travels along the corridor between the cultivation benches on the second lap. Here, the trolley is in second gear, the maximum speed is 0.25 m/s, and the loaded weight is 10 kg; the trolley sprays a straight section while applying the travel compensation control. The ROS function records the entire traveling and the trajectory of the trolley.

Fig. 11 Experimental path of the autonomous movement mode

3. Results and discussion

3.1. Results for the evaluation of the travel compensation control

The movement trajectories of the trolley for each experimental pattern (a)–(c) are shown in Fig. 12. The remote-control mode frequently deviated from the path because of the unstable posture caused by unevenness on the road surface. Furthermore, the autonomous-movement mode is point-to-point control, which is off the specified center. The autonomous-movement mode with travel compensation control demonstrated the highest accuracy in following the path and improved the straight-line performance. Comparing the RMS error, the remote-control mode was 0.10 m and the autonomous-movement mode was 0.11 m. In contrast, the mode with travel compensation control significantly reduced the error to 0.023 m. Moreover, the movement trajectories of the trolley for each reference path +0.15 m, 0m, and −0.15 m are shown in Fig. 13. The trolley followed all designated routes, validating the route specification function within the travel compensation control. The RMS errors for each path were 0.039 m for the +0.15 m path, 0.023 m for the 0m path, and 0.034 m for the −0.15 m path. This confirmed that the trolley followed all the target paths with high accuracy. These experimental results demonstrate the effectiveness of the travel compensation control.

Fig. 12 Trajectory of the path in each pattern

Fig. 13 Trajectory of the path for each reference route

3.2. Results for the evaluation of the human-following mode

Figure 14 shows the trolley and operator during the experiment. It was confirmed that the trolley followed the worker while maintaining a certain distance during work. Travel compensation control enables the trolley to travel along a set route regardless of the operator’s position in a straight section. In addition, the operator can harvest simulated crops in parallel with trolley operation. The trolley effectively followed the worker while maintaining a safe distance, allowing for hands-free operation and reducing the workload.

Fig. 14 The experiment in the human following mode

3.3. Results for evaluation of the autonomous-movement mode

Figure 15 shows the trajectory after the second lap, and Fig. 16 shows the trolley during the experiment.

Fig. 15 Trajectory of the autonomous movement mode

Fig. 16 The experiment in the autonomous movement mode

During spraying, the trolley was stabilized by the travel compensation control in the passage between the wall and the cultivation bench during the first lap. Next, we confirmed that the trolley had reached the charging station and started charging by the entering operation. After a certain period, the trolley left the charging station and began its second lap. During the second lap, the trolley was stable during spraying owing to the travel compensation in the passage between the cultivation benches. Finally, it was confirmed that the trolley reached the charging station and began charging. Regarding the travel accuracy in the spraying section, the error between the designated path and the actual position was within 0.1 m. Consequently, the autonomous trolley could perform unmanned spraying, enter and reenter the charging station, and travel in parallel with the spraying process while traveling in a straight section. Furthermore, we confirm that localized automatic spraying is also possible by designating specific positions, which effectively improves worker safety. Thus, we confirmed the automation of multiple agricultural operations over a long period, including recharging. The trolley successfully performed unmanned spraying, navigated to the charging station, and completed multiple agricultural operations autonomously, demonstrating its potential for long-term unmanned operations.

4. Conclusions

This study proposes an agricultural work support trolley to assist in greenhouse cultivation. A prototype agricultural support trolley was developed by converting a commercial trolley into a motorized trolley to provide support for multiple tasks. First, we equipped the trolley with a human-following mode. This mode supports tasks such as harvesting by following the worker. Second, it has an autonomous-movement mode for unmanned work such as spraying pesticides. In addition, we added a control system to maintain its posture and position between the benches. This improves its stability and makes it easier to operate.

Through evaluation experiments, we confirmed that the system could reduce harvesting work by enabling hands-free operation of the trolley in the human-following mode and automate pesticide spraying work by autonomous movement. A prototype demonstrated its effectiveness in reducing the burden of harvesting and automating pesticide spraying. Future work will include long-term operational testing in actual greenhouse environments, along with quantitative evaluations of task duration, work efficiency, and liquid adhesion, to further enhance practical applicability. In addition, the integration of an onboard camera is expected to add monitoring capabilities and contribute to further multifunctionality.

Acknowledgments

This work was supported by JSPS KAKENHI Grant Number JP21K05858.

Declaration of conflicting interests

The authors declare no conflicts of interest.

Notes

(URLs on references were accessed on 27 January 2026.)

References

•  Åstrand,  B. et al. 2005. A vision based row-following system for agricultural field machinery. Mechatronics. 15 (2): 251–269. https://doi.org/10.1016/j.mechatronics.2004.05.005
•  Bai,  Y. et al. 2023. Vision-based navigation and guidance for agricultural autonomous vehicles and robots: A review. Computers and Electronics in Agriculture. 205: 107584. https://doi.org/10.1016/j.compag.2022.107584
• e-Stat. 2022. Survey of agricultural management / Agricultural statistics by farm management type, 2022. Japan: the National Statistics Center. (In Japanese.) https://www.e-stat.go.jp/stat-search/files?stat_infid=000040168764
•  Fujinaga,  T. et al. 2021. Development of agricultural robot for work support in small-scale protected horticulture -Verification of map construction in different environments-. Proceedings of the 64th Japan Joint Automatic Control Conference. 272. (In Japanese.) https://doi.org/10.11511/jacc.64.0_272
•  Fukao,  T. 2021. Field robotics: Applications and fundamentals. Journal of Robotics and Mechatronics. 33 (6): 1216‒1222. https://doi.org/10.20965/jrm.2021.p1216
•  Han,  C. et al. 2023. Visual navigation and obstacle avoidance control for agricultural robots via LiDAR and camera. Remote Sensing. 15 (22): 5402. https://doi.org/10.3390/rs15225402
•  Hayashi,  S. et al. 2012. Study on an inside-approach harvesting robot for picking strawberries. JOURNAL of the JAPANESE SOCIETY of AGRICULTURAL MACHINERY. 74:(4): 325‒333. (In Japanese.) https://doi.org/10.11357/jsam.74.325
•  Imou,  K. 2017. Automatization and robotization of agricultural machinery. Journal of the Robotics Society of Japan. 35 (5): 376‒378. (In Japanese.) https://doi.org/10.7210/jrsj.35.376
•  Kim,  D. N. et al. 2015. Volatilisation of pesticides after application in vegetable greenhouses. Science of The Total Environment. 505: 670‒679. https://doi.org/10.1016/j.scitotenv.2014.10.036
•  Kurs,  A. et al. 2007. Wireless power transfer via strongly coupled magnetic resonances. Science. 317 (5834): 83‒86. https://doi.org/10.1126/science.1143254
• MAFF (Ministry of Agriculture, Forestry and Fisheries). 2023. Agricultural fatalities in 2021. Japan: MAFF. (In Japanese.) https://www.maff.go.jp/j/press/nousan/sizai/attach/pdf/230210-4.pdf
• MAFF. 2024a. Agricultural DX initiative 2.0. 3‒4. Japan: MAFF. (In Japanese.) https://www.maff.go.jp/j/kanbo/dx/attach/pdf/nougyou_dxkousou-167.pdf
• MAFF. 2024b. Summary of the annual report on food, agriculture and rural areas in Japan. 18. Japan: MAFF. (In Japanese.) https://www.maff.go.jp/j/wpaper/
•  Mizumoto,  M. 1995. 1. Product-sum-gravity method = fuzzy singleton-type reasoning method = simplified reasoning method. Proceedings of the Annual Conference of Biomedical Fuzzy Systems Association. 8: 17‒18. (In Japanese.) https://doi.org/10.24466/pacbfsa.8.0_17
•  Mousazadeh,  H. 2013. A technical review on navigation systems of agricultural autonomous off-road vehicles. Journal of Terramechanics. 50 (3): 211‒232. https://doi.org/10.1016/j.jterra.2013.03.004
•  Nagasaki,  Y. 2021. Current status and issues of machines and robots to reduce the burden of farm work. SYSTEMS, CONTROL AND INFORMATION. 65 (12): 467‒470. (In Japanese.) https://doi.org/10.11509/isciesci.65.12_467
• Open Robotics. 2016. navigation. USA: Open Robotics. http://wiki.ros.org/ja/navigation
• Open Robotics. 2019. gmapping. USA: Open Robotics. https://wiki.ros.org/gmapping
•  Przybyla,  M. 2017. The obstacle_detector package. https://github.com/tysik/obstacle_detector
•  Sasaki,  Y. et al. 2011. Detection of the intention from gestures of workers for a Kansei agri-robot. Agricultural Information Research. 20 (1): 13‒18. https://doi.org/10.3173/air.20.13
•  Ueda,  R. et al. 2004. Expansion resetting for recovery from fatal error in Monte Carlo localization - comparison with sensor resetting methods. Proceedings of the 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). 2481‒2486. https://doi.org/10.1109/IROS.2004.1389781
•  Xiong,  Y. et al. 2019. An autonomous strawberry-harvesting robot: Design, development, integration, and field evaluation. Journal of Field Robotics. 37 (2): 202‒224. https://doi.org/10.1002/rob.21889
•  Yorozu,  A. 2021. Demonstration and evaluation of agricultural robot supporting farmers. Proceedings of the 64th Japan Joint Automatic Control Conference. 269. (In Japanese.) https://doi.org/10.11511/jacc.64.0_269