1 Introduction

Automated guided vehicle (AGV) systems have been used as material transportation systems since the 1950s. According to the VDI 2510 standard guideline (as cited in [1]), an AGV is a “floor-supported, self-propelled means of transport which are controlled automatically and guided by a non-contact guidance system.” Currently, many industrial branches—such as Food and Beverage Industry and Electronics Industry—and non-industrial environments –such as warehouses, hospitals, and hotels—employ AGVs in their automation systems. Ullrich [1] presents the classification of the technological evolution of AGV systems in four eras. The sensors and sensing techniques employed are well demarcated from one period to another. The extensive use of contact-free sensors, such as magnetic and laser navigation, marked the Third Era (from 1990 to 2010, approx.). At the current one, the Fourth Era, changes will result in “new, low-cost, and intelligent sensor systems” [1]. Yet, [1] affirms the technologies of the third era are well established, presenting standard solutions. On the other hand, the use of cameras, radar, and ultrasound sensors are still under development.

The position/movement control of an AGV is a crucial point of its operation. The position control problem means the AGV to follow a path with minimized distance and orientation angle errors compared to the reference. Adequate AGV position control leads to greater movement accuracy and greater AGV lateral waving stability. Meeting some advantages such as work in smaller corridors, use more relaxed load and unload devices, the safe transport of delicate loads, among others. Furthermore, a key feature of such a control system is the sensor system. Choosing the AGV sensor is an essential but not trivial task, due to its cost and measurement evaluation, as well as determining the processing, filtering, or fusion method [2].

The present research aims to identify the sensors, and related sensing techniques, in machine vision applied to the AGV position control in the past 5 years published academic researches. In doing so, we focus on classifying the relations of AGV sensors and (a) the utilization area of the vehicle, (b) the adopted control strategy, and (c) the level of accuracy or sensibility required from the AGV. Due to the number of theoretical concepts and different technologies related to the selected primary studies (PSs), the paper does not aim to encompass an in-depth technical description. However, it discusses the technologies briefly, pointing out key references. At last, the present paper intends to draw an overview of the research field. To the authors’ knowledge, there is no other work examining the same subject.

For these reasons, the present paper proposes a systematic literature review (SLR) on the sensors and related techniques applied to the AGV positioning control problem from the past 5 years of literature. The systematic literature review is a scientific and formal method for recognizing, assessing, and interpreting the available PSs concerning a research topic [3]. Unlike ad hoc or narrative literature review, an SLR has a restricted research topic and review protocol. According to [4], the SLR is a qualitative review since it addresses “questions about the specific use of a technology” but does not intend to make quantitative comparisons. The present paper details the review protocol to allow reliability and reproducibility. The SLR identifies the sensors and sensing techniques employed in the problem domain by data extraction of the selected PSs, and a critical qualitative analysis of its results.

Despite the increasing number of patents concerning AGV technologies, the present work focuses on academic research and the defined period published results. So, we do not look for patent deposits and bases. It is a delimitation of our SLR. Another delimitation concerns the problem domain. Here, we differentiate position control from navigation techniques because both are also used in the AGV position control problem. Consider the classification of robot control architectures in deliberative or reactive [5]. The main characteristic of a deliberative architecture is the sense-plan-act paradigm. Such control anticipates the future system states, by sensing the environment, to decide the robot’s actions. On the other hand, a reactive architecture is a set of reaction rules that acts from the sensors reading. A reactive control does not represent the environment with a map, for instance, which the deliberative one demands. An example of a deliberative control of an AGV is a navigation system with localization and path planning modules. A line follower algorithm with a PID controller is an example of reactive control. Also, one can use both control paradigms in a hybrid control architecture. In that case, the reactive control belongs to a lower-layer regarding deliberative since the former receives commands from the latter. The SLR limits the problem domain to reactive or low-level control. It means that although the authors recognize that several pieces of research employ navigation techniques to the AGV position problem, in this paper, we address the reactive approaches.

We use an SLR software tool called StArt (State of the Art through Systematic Review) [6] to manage the references, assess their quality, and extract PS information. We used two comprehensive coverage databases in the SLR, Scopus, and Web of Science, finding 379 articles at first. After filtering and snowballing from the chosen papers, we selected 31 PSs for full-text reading and analysis. It provides an overview of the current research area state and proposes a research agenda.

The structure of the paper remaining is as follows: we present the SLR protocol and discuss its steps in the next section; Section 3 summarizes the results answering the research questions; finally, Section 4 provides the concluding remarks.

2 Systematic literature review protocol

Before starting the SLR, we verified if a similar review was already published. Using the search string [(“AGV” OR “Automated Guided Vehicle”) AND (“Systematic Review” OR “Systematic Literature Review”)] we have not found a matching result at Scopus, Web of Science, neither on Google Scholar. After no systematic review concerning the topic was found, we conduct the present SLR following the protocol steps Fig. 1 shows. Each process step has its outcome description.

Fig. 1
figure 1

Systematic literature review steps: processes and outcomes

Full size image

It leads to PS identification, selection, and research evidence analysis describing the methods, the search engines used, the inclusion and exclusion criteria, among other aspects. So, an SLR allows reproducibility avoiding researchers’ bias. According to [3], the most common reasons for undertaking an SLR are to summarize the existing evidence over a topic, identify research gaps, and provide a background to positioning new research activities concerning the topic. The present paper fits the first two reasons.

Still, [3] affirms an SLR has the advantage of providing information about a phenomenon from several different settings and empirical methods. It occurs despite a significantly more effort needed compared to traditional reviews one can see as a drawback. So, the SLR process helps to describe the current state of the AGV position sensors research topic.

For primary studies search and identification, we used the Scopus and Web of Science databases. The databases chosen have a wide scope, including several publishers, conferences, and journals in their search results. As the nomenclature to AGV is a little fuzzy in the literature, we use several synonyms and also search for the term mobile robots. Our target was to find PSs in AGV position control to extract the information about its applied sensors and sensing techniques. Table 1 shows the search strings used in each database with their particular terms and settings. It also shows the raw number of PSs returned from each search and the total number of entries. In Section 2.2, we discuss the study selection steps.

Table 1 Search strings used in each search engine
Full size table

We use StArt [6] to support the SLR processFootnote 1. StArt supports all SLR steps. The tool includes the protocol description, the study selection, managing, and pointing to duplicated studies, the quality assessment of the select studies, and its data extraction. Also, it provides data analysis and synthesis mechanisms by generating graphs such as word cloud, authors relationship, studies, and reference relationship, among other features.

2.1 Research questions

The research questions (RQ) motivate and guide the review process. In this work, we have a main RQ and connected sub-questions derived from the first. The sub-questions (sQ) help to answer the main one from three different points of view. Our main question is:

RQ:

Which sensors and sensing techniques are used in indoor AGV position control problems according to the past 5 years published researches, and which is their technological impact?

From the main question, we derive three sub-questions:

sQ1:

Is the sensor/sensing technique related to the AGV application area?

sQ2:

Is the sensor/sensing technique applied to the problem related to the control strategy and/or the AGV guide?

sQ3:

Is the sensor/sensing technique related to the required AGV accuracy/sensitivity level?

By answering sQ1, we intend to identify any sensor application recommendation—or rejection—related to the AGV purpose and identify a (possible) sensor classification by the vehicle employment. Such a question also measures how authors consider the appropriate sensor to the AGV operating conditions. With sQ2, we aim to understand how the sensor/sensing technique influences the position control strategy—if it influences. The interrogation covers the (possible) classification of sensors by controllers strategies, too. Finally, by responding sQ3, we propose to discuss if the authors consider AGV operating requisites when choosing the sensor/sensing technique. Here position accuracy means the AGV offers a smaller position error compared to its reference (virtual landmarks, magnetic tape, colored stripe, among others). The sensitivity level means the AGV load oscillation tolerance due to its fragility. As a qualitative analysis, the sub-question answering does not intend to compare accuracy or sensitivity among PSs. It aims to investigate how PSs represent the control system requisites—if they do—and how these requisites influence the sensor choosing or the sensing technique adjustment. A comparison of accuracy improvement within a PS is possible if the authors compare their proposed approach to others or compare two types of sensors. The answer to the main question provides an overview of the research area’s current state.

2.2 Studies selection: steps and criteria

To avoid bias, we defined the inclusion and exclusion criteria before the studies screening. It prevents the studies from influencing the criteria definition. The study selection steps are, following the execution step identification in Fig. 1: (a) studies identification; (b) checking for duplicates; (c) selecting studies according to their titles; (d) selecting studies according to their keywords and abstracts; (e) selecting the studies after a full-text reading; (f) snowballing from the previous selected studies—selecting studies according to their titles; and (g), (h) applying filters until the final selection. To include relevant studies and exclude non-relevant ones, an SLR process needs clear inclusion/exclusion criteria. We established six inclusion and six exclusion criteria, shown in Table 2.

Table 2 Studies inclusion and exclusion criteria used in the SLR
Full size table

To include a study at stages (c) and (d), it must satisfy criteria I1, I2, and, at least, another one from I3 to I6. The same applies to studies selected by snowballing, in stage (g). The final inclusion occurs if the study meets I7 after the full-text reading, in step (e) and (h)—for the snowballing selection. Similarly, in (c) and (d), any of the criteria from E1 to E5 exclude a study, and, in (e), E6 does. The same applies to studies selected by snowballing, in stages (g) and (h).

The SLR process started with 379 studies after the (a) studies search and identification. In step (b), we eliminate the duplicates using StArt duplicate checking. It results in 298 studies. By reading studies titles and first applying the inclusion and elimination criteria, step (c), we reduced from 298 to 180 studies. Next, from step (d), we achieved 86 studies. After a full-text reading, step (e), we excluded 57 studies, resulting in the final selected group of 29 PSs. From the selected papers, a snowballing reading reached 18 studies. Applying the criteria, we end up with two studies after full-text reading. So, the present SLR analyzes 31 PSs.

2.3 Data extraction and quality evaluation

From the selected PSs, we extracted the following features: date of publication, country of the researchers, which position sensors the study employs, which technique the study uses to process the sensor data, the type of AGV guidance, the application area of the AGV, the control strategy used, if the PSs compare the proposed control with another one, how the proposed sensing technique impacts the AGV position accuracy, and if the study presents a sensor fusion technique. Table 3 relates each extracted feature with a sub-question. Despite our interest in the experimental part to analyze the sensor use, we find relevant studies regarding sensor employment, but with a simulated AGV position response.

Table 3 Extracted features from the selected studies and their relation to the research questions
Full size table

To assess the quality of the selected PSs, we developed a quality assessment form. It aims to assist in data analysis and synthesis. Table 4 shows the quality assessments of the selected studies. It adapts the quality assessment items and scoring from [7] and was construct based on the systematic review literature [3, 4]. The quality form has five assessments. Each one has a list of possible answers with specific values, as shown in Table 4. A primary study (PS) quality assessment score varies from 0 to 16 points by summing up the quality assessment answers’ values. Concerning the research sub-questions and main question, QA1 and QA2 relate to sQ1, QA3 and QA4 relate to sQ2, and QA5 relates to sQ3. Table 4 shows the number of PSs of each QA item in the #PS column.

Table 4 Quality assessment form, adapted from [7]
Full size table

2.4 Threats to SLR validity

We identify four main threats to SLR validity. The first one is the possibility of missing meaningful PSs because of a biased search string. To mitigate the threat, we used a control groupFootnote 2 to refine the search string. Also, we intended to include several synonyms terms for AGV. Despite our effort, the SLR may miss relevant works. The next recognized threat is researchers’ bias. The authors have no conflict of interest in this research. Despite the authorship of a selected study, the bias threat is minimal. Another threat refers to the data extraction method. We developed the data extraction form according to the SLR aims. Not all the information in PSs were obvious to fill the form. The authors discussed such cases to mitigate the validity threat. Lastly, we admit the quality assessment threat. Since we developed the quality assessment form to assist the data extraction and analysis step, it may bias the study quality score because of our focus, but it does not exclude any work.

3 Results and analysis

This section presents the results according to the data extraction and quality assessment of the PSs. Next, we present a brief SLR overview. Also, the present section discusses each research sub-question to reach the main research question. It culminates in a research agenda proposal/trend presentation/research gap presentation.

The selected PSs are listed, per search session (SS) and year, in Table 5. According to Fig. 2, most of the selected papers were published in the past 2.5 years—considering papers published since 2018—indicating the research area still attracts researchers.

Fig. 2
figure 2

Feature F1: Distribution of publications over the past 5 years

Full size image
Table 5 Selected primary studies references according the search session of Table 1, divided by year
Full size table

Following the PSs’ features extraction, Fig. 3 presents the papers’ regional distribution. China and Korea are the more prolific countries, which may be related to their manufacturing automation level and the number of manufacturing sites.

Fig. 3
figure 3

Feature F2: Regional distribution of publications over the SLR analyzed period

Full size image

Regarding the study type classification (F3), three PSs are simulation only studies: [22, 30, 36]. Despite presenting relevant use of a positioning sensor in AGVs, those studies do not demonstrate the application in a real environment, displaying a numerical simulation of the proposed position control. Another five papers—[11, 18, 31, 34, 37]—present experimental results and a theoretical development too, which we classify as mixed studies. For instance, [11] propose a fault detection algorithm for trajectory tracking based on different sensors and their respective position estimation modules. The algorithm uses an extended Kalman filter to detect a fault in the estimation modules. By isolating a faulty module, it protects the AGV trajectory response. Zhang and Huo [37] propose an integrated mathematical model of the inertial navigation system and the vision navigation system to apply a Kalman filter to estimate the AGV position errors. The other 23 studies exhibit and discuss experimental results without further theoretical developments. Table 4 shows the number of PSs per quality assessment item—#PS column—and Table 6 shows the score of each paper. In the next sections, we discuss each quality assessment related to the respective sub-question.

Table 6 Quality assessment results
Full size table

Finally, to summarize the data extraction, we present Table 7. It shows each feature extracted from the PS according to the ones listed in Table 3.

Table 7 Data extraction summarizing: extracted features F3 to F10 of selected PSs
Full size table

For simplicity and to help an overall vision, we aim to classify the extracted information by similarity. For example, regarding F6, the papers which use the camera information as the primary measurement of the AGV position have the same label “image processing,” despite different approaches. The same occurs with F8 and F9, in particular. The following sections present the data extraction results discussing each sub-question and the main research question.

3.1 Sub-question 1

This research sub-question aims to recognize a relationship between the sensor/sensing technique and the AGV application area. With the PS data extraction, Table 7 shows the primary position sensor (F4) and the sensing technique (F5) used in the PS, the AGV guidance technology (F6), and the application area (F7), following the PS description. Feature F6 is closely related to F4 and F5. Alternatively, the guidance type drives the sensor’s choosing, or the opposite occurs. Additionally, QA1 and QA2 results, in Table 4, assist the sQ1 answering.

From the F4 extraction results, almost half of the selected PS uses a camera as a position sensor, i.e., take environment measures from image processing. Some papers use external cameras. [12] uses an external optical measurement system to control an AGV position and synchronize its movement to other equipment in a continuous conveyor system. The previous research also uses a colored stripe and an on-board AGV camera, possibly using local position control despite no further discussion. [32] identifies the AGV position in a warehouse by using two external cameras to recognize a reference mark at AGV top. [26] uses a stereo camera to detect QR Codes and measure the AGV position.

The remaining PSs are divided into (i) approaches using the camera to identify a colored stripe to measure the AGV distance and orientation angle, and (ii) approaches using the camera to identify specific image codes to measure the AGV trajectory deviation and its orientation angle.

The PSs of the first group have two possible camera locations: the camera heading downwards with the image sensor parallel to the ground [15, 21, 24, 29, 33, 35], or the camera heading forward, pointing to the surface ahead of the AGV [36, 38]. Despite having a restricted field of view, the downward camera has allow the AGV designer to embed the sensor in the AGV structure. It reduces the external lighting influence over camera readings. Commonly, the sensor structure, in this case, has a separate source of light [21, 33]. Those PSs use similar algorithms to process the image. [15, 21, 29, 35] use the image center as the reference, both vertically as horizontally. Its algorithms measure the distance error by identifying the stripe’s center pixel and calculating its distance from the vertical reference. The same stands for the AGV angle, comparing a line angle (formed for two [21] or three stripe pixels [15, 29], or extracted from stripe edge detection [35]) with the reference. In addition, [21] discuss in detail the used image processing algorithm, presenting the algorithm, equations, the conversion from pixel to distance, and the morphological operations used on image processing. Differing from the mentioned algorithms, [24] present a type-II fuzzy-based system to determine vehicle displacement. It calculates the pixel density to estimate the guide’s location on the image frame, divided into regions. Those regions are the fuzzy membership functions. The method allows for varying the number of regions to increase accuracy.

About the other camera location, as [36, 38] show how non-uniform illumination, for example, affects the image processing. Moreover, such applications must consider the image plane distortion caused by the camera position and the field of view distortion [38]. As an advantage, the camera positioning extends the field of view, allowing the sensor system to recognize in advance the path [36].

About the methods (ii), one can observe that most of the PSs use other sensors together with the camera [27, 31, 37], such as IMU (Inertial Measurement Unit) and laser scanners, to measure the AGV deviation between the landmarks. It also enables a larger space between the landmarks on the floor. Yet, [10] presents another similar approach using Bluetooth landmarks to guide an AGV. The use of such landmarks enables a distance estimation from the beacon level.

The other half of studies use standard solutions, as [1] points out. The magnetic sensor is used in seven PSs. Although four of them use the magnetic tape as the AGV guidance technology, three use magnetic nails. The last use is similar to the use of landmarks discussed in the previous paragraph. Its use also indicates the use of other sensors the measure the AGV position between the magnetic nails, which are used to correct the AGV orientation. A single PS uses an optical sensor, similar to the magnetic one, with discrete measures of distance and orientation angle of a colored stripe on the floor [8].

The remainder PS use laser rangefinder, laser scanner, or lidar. We divide the PS that use laser scanners and lidar into approaches using laser reflectors [11, 25] and those using a virtual path [9, 14, 20], traced through environment mapping. The first group uses an established method, but the main disadvantages are the need to calibrate the position of the reflectors and its inflexibility to changes due to the first drawback. The second group faces a challenge from the AGV systems Fourth Era of making vehicles autonomous [1]. Pratama et al. [11] use a SLAM (simultaneous localization and mapping) algorithm to locate the landmarks and an EKF (extended Kalman filter) to estimate the AGV position as it moves and updates the encoders odometry value. SLAM is, according to [39], “the process by which a mobile robot can build a map of the environment and, at the same time, use this map to compute its location.” Additionally, [39] describes SLAM’s fundamental algorithms, such as probabilistic SLAM, EKF-SLAM, used in [11], and Rao-Blackwellized Filter. Bui [14] used a laser-based localization sensor but did not disclose the localization algorithm. Ye and Zhou [20] propose a double-layer EKF-SLAM. The inner EKF layer predicts the AGV’s position and uses an electronic compass to correct the AGV’s pose, while the outer layer takes the output of the inner layer to correct the AGV’s pose regarding the environment map using the laser scanner measurements. The work of [28] uses laser rangefinders for a wall following strategy. Although functional, the strategy has application constraints.

As for F6, although it points to more commonly used techniques, such as Kalman filters and their derivatives, there is no clear association denoting an application recommendation.

In conclusion, from the selected studies, we cannot set a relation between the application area of AGV (F8) and the sensors used in the selected literature. PSs do not discuss the influence of the AGV working environment on sensor performance nor in sensing techniques. This distance from the actual application can be related to (i) the use of AGV prototypes and some solutions still only simulated, as measured by QA1 and (ii) the number of works presenting solutions for AGVs in controlled environments, as QA2 assessed.

3.2 Sub-question 2

Here we separate control strategies from navigation techniques, as delimited in Section 1. The selected studies have a low-level position control. Some studies apply a hybrid control architecture, meaning a high-level control that plans AGV’s action, such as path planning modules producing the desired path. PSs [10, 13, 20, 26, 37] using navigation algorithms do not present error correction methods as the ones using closed-loop control systems do. The papers [15, 16, 19, 27, 28, 32, 35] do not comment on the control strategy used, despite presenting the methods of measuring the AGV position and showing experimental results. We differentiate these two groups of works because the first group allows the reproduction of the experimental results, even if in part.

Among the control techniques (F8) employed are the classic PID control, Fuzzy controllers in different configurations, neural networks, and other feedback control strategies, as shown in Table 7. One can observe that, although there is no direct relation between control strategy and sensors, more sophisticated techniques (as adaptive neural networks and backstepping control) of controllers involve a more sophisticated sensing technique. PSs using PID make a simple correction of distance and/or orientation angle. If the control takes both variables, there is a simple weighting in the control law error input [21]. AGVs are non-trivial and non-linear dynamic systems. However, few studies use adaptive or non-linear control strategies that account for the inherent variation of vehicle load or possible speed change.

Another essential point to the sub-question is that most studies do not compare the proposed control strategies (F9). It does not enable a quantitative analysis of the proposed control technique and its influence on the AGV position accuracy. Despite lacking a comparison, most studies describe the used control strategies in a greater or lesser degree of detail. If a PS does not describe the controller, it refers to well-known techniques, as the QA3 results show. Likewise, most papers describe the sensing technique used, while eight papers do not address it. We assume those eight PSs do not detail the technique because (i) the used sensor gives a direct measurement, not requiring high processing or (ii) it was not the focus of the work.

After all, we cannot perceive a pattern between the sensors and the selected works’ control strategies. However, the SLR shows the need to expand adaptive controllers’ use to the AGV position control problem. Another detail is the vast majority of papers (29) do not apply a load variation on the vehicle, except [14] and [15]. That is, the authors do not test the AGV at all points of operation. The load change can even influence the sensors since it causes a drastic change in vehicle dynamics. Furthermore, the work lacks that object of investigation.

3.3 Sub-question 3

To answer this question, column F10 of Table 7 has a direct relationship with F9 data. Only by comparing control techniques we can measure a quantitative position accuracy improvement. The studies that present a qualitative improvement present a comparison with the application or not of the proposed sensing technique [28] or the difference in response with different distances between two landmarks [31]. Although the paper does not present experimental position control results to compare, [38] presented an improvement in the track recognition using the proposed image processing technique. Besides, [30] displayed improvements solely in the controller simulated results, excluding the sensor’s influence in the control loop. Furthermore, [21] discuss how the sensor parameter tuning impacts the position control accuracy. The other PSs that demonstrate accuracy improvements compare their control techniques with other techniques [9, 13, 16, 20, 24, 34] or compare the effect of the use of more or fewer sensors in the response of the position control of AGV [11].

Feature F11 is about sensor data fusion. About one-third of the works use some technique of sensor fusion. Using two heavy load tractor AGVs, each one with a camera, [18] used image processing to fuse sensors data to plan and correct the trajectory of the set. Yan et al. [22] applied a fuzzy inference system to fuse the data from the gyroscope and the magnetic sensor to correct the AGV orientation angle. The remaining PSs utilize the Kalman filter or the extended Kalman filter as a sensor fusion technique. Despite having another embedded sensor in the AGV, [33] did not employ a sensor fusion technique nor indicate using more than one sensor to measure the AGV orientation angle and distance deviation.

As not all studies compare their results with other techniques, it is impossible to relate the application of a sensor fusion technique to the accuracy of the AGV position control. Of 10 studies, [9, 20] presented fusion techniques comparison, [11] used more than two types of sensor data in the fusion method, and the other eight apply sensory fusion with two types of sensors.

All things considered, the selected PSs do not discuss the accuracy of an AGV position control, neither its relationship with the vehicle’s sensor system. This lack is further reinforced regarding the number of researches that describe the AGV operating requirements, such as maximum trajectory deviation tolerance. Only seven PSs outline the AGV operation requirements: [12, 15, 19, 24, 25, 27, 30], according to the QA2 results. Moreover, the other studies do not represent or disclose the AGV operating requisites. Therefore, they do not discuss how the sensor system attends system requirements, such as measurement update frequency, resolution, and noise rate. For instance, [21] presented how the camera resolution affects the total processing time of a vision-based position sensor, which impacts the control loop and the controller adjustment.

3.4 Main research question

The SLR presents an overview of the selected primary studies that help answer the first part of the research question (RQ). Table 7 lists the sensors and sensing techniques to answer the question. From this perspective, we realize that standard sensor technologies are still targets of new research, with new settings and new algorithm proposals. Also, it reveals the growth in the use of machine vision solutions. As for the techniques, qualitatively, we can analyze that there is space for research in the sensor fusion research area, intending to increase the accuracy of the position control of an AGV.

On the technological impact of the reported sensors and sensing techniques, we make a qualitative analysis. The number of researches on machine vision solutions meets [1] affirmation the changes of the Fourth Era of AGV Systems go through the use of new, low-cost, and intelligent sensors. Moreover, it is a challenge for researchers in this area to achieve standard solutions adopted by the industry. Likewise, established technologies can be improved with the use of other sensors and new sensing techniques.

3.5 Research agenda proposal

This section aims to propose a research agenda for future investigation on sensors used in AGV position control from the discussion so far. Other research areas and applications in AGVs are not the focus of this proposal, such as the vehicle’s safety sensors and its surrounding environment, or sensors used in outdoor environments. Based on the SLR findings, we list the following points which future researches should address:

To concern about the AGV application area and its influence on the position sensor response

From the read PSs, few discuss the environment noise influence on sensors measurements. An industrial site may impact a sensor measurement electromagnetic noises, with noise from the dirty (dust on suspension, dirty floor), from the indoor lighting (light variation caused by the manufacturing process), among others. For example, [40] listed several influences on laser scanners, being the surface reflectivity, the temperature, the presence of dust, or steam on the measured atmosphere. So, future research should concern and address these issues.

To include experiments exploring the load range of the vehicle

The load variation impacts the vehicle dynamics. From the controller perspective, this change can degrade the position accuracy performance and even lead the system to instability. Since an AGV purpose is to carry loads, any system proposal shall investigate its performance at all operation points of the system. Affecting vehicle dynamics may also impact the sensor measurements. For example, an unbalanced load may increase vehicle vibration. In this case, it causes a noisier laser scanner reading or a blurred image capturing, for example. With a heavier load, the wheel friction grows, which may impact encoders reading. So, AGV experiments to verify a controller or a sensor/sensing technique performance should cover the vehicle load range.

On the position controller proposition and the use of a sensor, discuss the AGV position accuracy requirements

Besides the description of the application environment, the investigation shall include the controller’s requirements or the sensor. A proper problem definition brings its demands. It leads to the next point.

To define AGV performance metrics

The AGV performance metrics definition enables a clearer comparison among different platforms and applications. Despite the effort in describing AGV System’s performance metrics [41, 42] and navigation performance metrics [43, 44], in the literature, we found no procedure or method defining the performance metrics for guided vehicles. So, the controller performance or the sensor system influence on the AGV position control cannot be compared to another approach without a reimplementation, which mischaracterizes the latter.

Research on machine vision sensors with a focus on standard solution achievement

As identified by [1], the use of machine vision sensors is still under development. The research is indispensable for the development of the sensors to meet regulatory standards and market acceptance.

Research on intelligent sensors focused on connecting already installed AGV systems to the Industry 4.0

Despite not being pointed out by the selected PSs, a future research path is studying intelligent sensors focused on retrofitting old vehicles without the need to change the infrastructure of the site.

4 Conclusion

The paper presented a systematic literature review conducted to identify the sensors/sensing techniques in machine vision applied to the AGV position control in the past 5 years of published research. The AGV position sensor is directly linked to the desired measurement from the vehicle. This choice influences AGV position accuracy, stability, and performance. Starting with 379 papers, we selected 31 primary studies following the defined protocol. The paper describes the method so the SLR results can be reproduced, updated, and compared with future reviews.

The main paper contributions are (i) the SLR results and analysis and (ii) the research agenda proposition. The SLR results and analysis show an overview of the technologies and techniques used in published researches. We did not find a direct relation to answer each sub-questions, i.e., from the selected studies analysis, we cannot explicitly relate the sensor used to the AGV application area (sQ1). The same stands for the control strategy used (sQ2) or the required AGV accuracy/sensitivity level (sQ3). However, the results enlighten the subject to future researchers, pointing out references with diverse approaches. Also, we can highlight the systematic review method presentation and its application to a research area not explored before, as a secondary contribution. From the sQ1 investigation, we observe most works use of AGV prototypes in a controlled environment, e.g., the laboratory, or even simulated AGVs. It does not represent the real-world application. The studies also do not discuss the environmental influences, such as noises and uncertainties, on the AGV sensor. Concerning the sensor association to the control strategy, the studies apply a range from the classic control theory to adaptive strategies.

Nevertheless, most works do not test the AGV at all points of operation, which requires controller robustness in terms of its performance facing the system dynamics change. The load change can even influence the sensors since leads to a drastic change in vehicle dynamics. Besides, most of the PSs lack that object of investigation. Regarding the third sub-question, the SLR shows, in general, papers do not discuss the accuracy of an AGV position control, neither its relationship with the vehicle’s sensor system. It means that authors do not discuss how the system requirements influence the sensor system choosing and the control design.

Since half of the primary studies use machine vision techniques, the RQ’s answer stresses new machine vision sensors’ value to the AGV position control problem. The research area overview also shows that standard and well-accepted sensor solutions are used with other sensors to improve response. Despite the effort for new solutions, we indicate that researchers must intensely discuss the AGV environment’s influence, the load variation, and the vehicle operational requirements when proposing a sensor system or a control system to an AGV.