2. Fachtagung TestRig - September 2024 17 Innovative Test Field Approach for Agricultural Applications Martin de Fries, M.Sc. Cologne University of Applied Sciences, Cologne, Germany Department of Electrical Engineering, Uppsala University, Uppsala, Sweden Marcus Irmer, M.Sc. Cologne University of Applied Sciences, Cologne, Germany Department of Electrical Engineering, Uppsala University, Uppsala, Sweden Karin Thomas, PhD Department of Electrical Engineering, Uppsala University, Uppsala, Sweden René Degen, PhD Cologne University of Applied Sciences, Cologne, Germany Department of Electrical Engineering, Uppsala University, Uppsala, Sweden Abstract In agricultural technology, the importance of near-operational complete vehicle tests is constantly increasing due to the growing degree of automation. This is because the interaction of attachments, tractors and drivers, sometimes with several assistance and automation functions, requires significantly more test kilometers in the field. However, additional test kilometers mean additional time that needs to be invested. Testing in the agricultural field is more dependent on weather conditions and seasons than in other sectors. Also, defined scenarios with certain crops are only available temporarily depending on the harvest year. One solution are virtual test fields: A principle, which has already become firmly established in automotive development. They offer fully reproducible test scenarios, temporal independence and test parameters that can be varied at will, while being significantly more cost and time efficient. In this work, the needs of the industry and the state of the art of highly realistic virtual test fields for agricultural applications will be discussed. Based on this, a modular approach for implementation will be presented. Subsequently, individual modules and their interfaces will be introduced. Rounded off by the presentation of practical application examples from the field of environment detection, the article concludes with insights into future development perspectives and functional enhancements. Keywords virtual test field - agricultural vehicles - sensor models - testing - automation - assistance systems - precision farming 1. Introduction Current times are characterized by enormous technological trends. For example, the enormous potential of digitalization is being used to simplify people’s lives and make it safer. A good example of this is the introduction of highly developed driver assistance functions in modern cars. The consumer demand will propel the growth of vehicles sales with autonomous-driving systems and advanced driver assistance systems (ADAS) from today. Meanwhile, the share of functions with high levels of automation will increase. [1] One of the critical enablers of the progress of these functions is the use of virtual development and testing methods. These innovative approaches offer substantial potential by allowing developers to simulate and test complex scenarios that would be impractical or impossible to recreate consistently in the real world. Virtual environments provide a controlled setting where various parameters can be manipulated to assess system performance under a wide range of conditions, significantly accelerating the development process and improving the robustness of the final product. In the context of this work, these methods are summarized under the term virtual test field (VTF). To avoid ambiguity, this term is defined here as a co-simulation environment for model-based system development [2]. Its core is a high-resolution, three-dimensional and geometric model of an operational environment of the vehicle system to be tested as well as the geometric models of the vehicle itself and other elements of the environment. The VTF can also contain other modules such as sensor models, vehicle dynamics models, human-machine interfaces, interfaces for hardware integration and other modules. Such a tool is suitable for prototyping and testing software-based functions of mechatronic systems in early development phases. It supports “virtual testing”, which is described as a development phase for mechatronic systems in the V-model according to [3]. Fig. 1 shows the VTF in the context of the V-model development process with premature branch for virtual testing. VTF are now integral to many industries, notably the automotive sector, where they are employed to refine driver assistance systems and automated driving Dieser Beitrag wurde durch den Programmauschuss peer reviewed 18 2. Fachtagung TestRig - September 2024 Innovative Test Field Approach for Agricultural Applications Fig. 1: A VTF supports the premature V-Model branch functions. Leading automotive manufacturers and technology companies are at the forefront of this innovation, combining simulation tools such as X-inthe-Loop (XiL) systems with virtual environments. Nowadays, VTF are of increasing importance for the automotive industry. According to a 2022 study, the market size for autonomous vehicle simulation tools is expected to have annual growth of 13.4% during the forecast period from 2022 to 2031. [4] The automation of processes is also playing an increasingly important role in agriculture. Smart and mobile farm machinery based on innovative sensor and computer vision technology could create $50 billion to $60 billion of additional market value by 2030 [5]. Furthermore, such agricultural vehicles and implements have the advantage that a significant proportion of their tasks can be relatively easily automated. One reason for this is that many tasks are repetitive. These tasks mainly include manipulative processes such as mowing, hoeing or harvesting, which are often carried out by dedicated implements and mobile machinery. During these processes, the tractor/ implement combination always travels the same routes. This is in contrast to the use of autonomous cars in road traffic, where the interaction of many road users and the use of complex routes result in constantly changing, complicated scenarios. Driving speeds are also significantly higher there. As a result, different requirements apply to assistance functions for off-road agricultural operations, which leads to regulations and development processes of varying complexity. In the agricultural machinery industry, the use of VTFs in the development of assistance functions is not yet common practice. However, such methods are the subject of current research and are described as a key future tool in the scientific field [6]. Further examples from research are described below. The objective of this paper is to describe the application of VTF in the agricultural technology sector, where the increasing automation of mobile agricultural machinery demands extensive testing. Near-operational complete vehicle tests are becoming more critical as the interaction between implements, tractors and operators, often involving multiple assistance and automation functions, requires extensive field testing. However, the reliance on physical testing is constrained by weather conditions, seasonal availability of crops and the significant time investment required to accumulate sufficient test kilometers. VTF present a promising solution by offering reproducible scenarios, temporal independence and flexible test parameters, all while reducing time and financial costs. This paper will discuss the needs of the agricultural industry and the current state of VTF tailored for agricultural applications. It will introduce a modular approach for implementing a VTF, detailing individual modules and their respective interfaces. Application examples will be presented, demonstrating the effectiveness of this approach. 2. State of the Art of Virtual Test Fields in Agriculture In order to formulate the desired properties of VTF for automation functions for mobile agricultural machinery below, properties and deficits of the state of the art must also be included. The literature research process used is based on the methodology of Integrative Literature Research as described in [7]. Its criteria for inclusion and exclusion are documented in Tab. 1 beginning with the definition of the search terms. These terms cover the field of simulation, automated vehicles and agricultural applications. As virtual development environments are often developed by foundations or companies and described on the associated websites, a browser-based web search is conducted. In addition to this, scientific journal publications were analyzed to investigate the use cases of such tools in research. The databases used, the publication period of the articles and further criteria are documented. Tab. 1: Methodical Criteria of the Literature Research 2. Fachtagung TestRig - September 2024 19 Innovative Test Field Approach for Agricultural Applications The evaluation begins with an analysis of the open source tools. These are each briefly described and analyzed with reference to the scientific sources with regard to their functions as a development tool, the structure of the software, the modeling of the vehicle system and its functions and the modeling of the virtual environment. The analysis of the open source tools is followed by an evaluation of the commercial solutions. For better clarity or a more concise review of the state of the art, Fig. 2 should be consulted, which analyzes the tools described below and their agricultural use cases. Fig. 2: Research Results for Tools and Use Cases 2.1 Open Source Solutions Among the open source solutions, the CARLA simulator should be mentioned. CARLA is a simulation platform for the development, testing and validation of autonomous vehicle systems [8]. It is frequently used in research and industry to develop and test ADAS and autonomous vehicle technologies. The platform is based on the Epic Games Unreal Engine 4 [9]. With the “Town 7” package, CARLA has so far only marginally addressed the simulation of agricultural land. The company AVL has integrated CARLA into its toolchain for the simulation of off-road vehicles and also mentions agricultural machinery as a use case [10]. CAR- LA is also used in research, for example in [11], which aims to create a realistic simulation for agricultural robots. CAR- LA offers advanced features such as hardware integration, data generation, operator integration and scenario configuration. The platform supports flexible configurations of virtual environments and adheres to the ASAM standards for designing road traffic scenarios [12]. Interfaces to the Robot Operating System (ROS) are available and the software can be extended by programming in C++, allowing interfaces to be created to other software for co-simulation, for example. The modeling of the vehicle and its functions in CARLA includes sensor models for LIDAR, RADAR and cameras, some of which provide realistic raw sensor data or ground truth sensor data. Ground truth refers to sensor data that is not generated by a physically-evident emulation of the sensor principle, but results from a query of correct and precise data points in the virtual environment. The simulation of physical processes, such as driving dynamics, is carried out by the Physics Engine of the Unreal Engine and can therefore not be defined by the user. The virtual environment in CARLA is characterized by high quality rendering based on Unreal Engine 4, although Unreal Engine 5 offers more advanced rendering options. Another open source solution is the Gazebo Simulator [13]. Gazebo is an open source simulation platform that provides virtual environments for modeling and testing robots and autonomous systems. It supports a variety of sensors and actuators and is often used in combination with ROS [14]. Gazebo enables the simulation of complex scenarios in real time. It is often used in the research and pre-development of robotics and automation technology. Clearpath Robotics uses Gazebo to simulate its robots in agricultural environments [15]. The agricultural machinery manufacturer John Deere, which bought a robotics company in 2017, also uses ROS [16]. In scientific studies, Gazebo is used to simulate crop yields and autonomous navigation in agricultural environments [17-19]. Gazebo offers features such as cloud integration for flexible use of maps and elements. Dynamic loading or discarding of environment elements is used to increase the performance of the simulations. Gazebo can also be flexibly extended via plugins. The software supports both console-based and graphical interfaces as well as web interfaces. The physics simulation is carried out using its own physics engine, which is described as more precise than that of game engines. The modeling of vehicles, robots and their functions in Gazebo includes models for sensors for environment detection as well as mechanical sensor models such as inertial measurement units. Noise processes are included in the signals of the sensor models to map realization effects, a strategy to make ground truth sensor data more realistic. The graphical representation of the environment appears detailed, but not as advanced as game engines such as the Unreal Engine. Blender is a marginal phenomenon [20]. This tool is primarily an open source software developed for 3D modeling, animation, rendering and video editing. However, the software also offers basic functions for physics simulation and motion simulation. Blender has not been developed for the simulation of vehicles and therefore does not reach the level of CARLA or Gazebo. It is also not specifically designed for real-time simulations. However, it is ideal for high-quality graphical modeling of virtual environments and their elements, which makes it particularly suitable for generating image data for camera sensor technology. The software can be controlled and manipulated via a Python API, but there is no indication on the integration other tools for simulation. Despite these limitations, Blender is used in agricultural vehicle research, for example in a study on the simulation of crop sensors [21]. No indication for the industrial application of Blender in the simulation of agricultural vehicle functions was found in this research. 20 2. Fachtagung TestRig - September 2024 Innovative Test Field Approach for Agricultural Applications 2.2 Commercial Solutions In the context of commercial tools, the company dSPACE is one to be mentioned [22]. dSPACE develops software and hardware for real-time simulation and the development of embedded systems. The company is strongly active in the automotive sector and offers a range of tools such as Aurelion [23] for sensor simulation and VEOS [24] for the simulation of dynamic models and ECUs, which together form holistic simulation solutions in the sense of VTF. Besides automotive companies, dSPACE also lists agricultural machinery manufacturers in its customer references and publishes concepts for agricultural application examples of their tools [25]. dSPACE offers extensive options for integrating hardware using XiL methods and for rapid prototyping. The company also develops special real-time hardware that is designed to meet the needs of automated vehicles. The scenario design is based on standards according to [12] and is supported by the dedicated configuration tools as Model- Desk [26]. The Aurelion Manager tool makes it easy to equip vehicle models with individual sensor configurations. Several software components work together in a virtual test and development environment based on dSPA- CE. Aurelion offers sensor simulations for RADAR, LIDAR and camera, which can generate realistic raw data as well as ground truth data. Various models for vehicle dynamics and other mechatronic systems are available for modeling vehicle systems and functions. The quality of the environment graphics appears similar to the representations in CARLA or Unreal Engine 4. The environment can also be automatically created and individualized based on user specifications. It is unclear whether and which physics engine will be used. IPG Automotive also develops commercial simulation solutions for various road vehicles [27]. In addition to vehicle dynamics simulations, the product portfolio also includes the simulation of assistance systems and automated driving functions. It also enables the integration of hardware into simulations for testing purposes. Use cases for the software are also shown an agricultural context [28]. Furthermore, applications from current research show joint use of IPG products with AgriSI from Soluzioni Ingegneria for the simulation of vehicle dynamics [6, 29]. IPG offers extensive possibilities for integrating hardware into the simulation, including real-time computers for sensor data calculation and a variety of XiL methods, such as vehicle or driver-in-the-loop simulations. Scenarios are set up in accordance with ASAM standards, manually or automatically using a 3D library and other databases. The environment for virtual testing is built around the CarMaker product, which is supplemented by visualization, configuration options, data flow structures and model libraries for automated driving. IPG offers comprehensive model libraries for the simulation of vehicle dynamics and powertrain, including exemplary models for agricultural tractors and implements in cooperation with [29]. Sensors such as RADAR, LiDAR and camera can also be simulated and ground truth data can be generated. The graphical representation of the environment is on a par with the visualizations from CARLA or dSPA- CE. The in-house engine Movie NX is used for rendering. Which physics engine is used is not described. The AgriSI software from Soluzioni Ingegneria is used to simulate vehicle functions and for driving simulation [29]. One aim is to train drivers in the use of precision farming functions through simulation and to develop suitable human-machine interfaces. There is no specific information on its use in industry, but the software is used scientifically, for example in [6] on co-simulation for precision farming applications. AgriSI supports driver-inthe-loop and further XiL methods. According to [29], the simulation of the vehicle system and its functions does not include sensor modeling. It works in combination with the model-based dynamics simulation from IPG Automotive. Communication between the vehicle and attachment can also be mapped using various protocols. The ISO- BUS protocol commonly applied in agriculture can also be used [30]. Scenarios are based on georeferenced maps, and routes or trajectories based on GPS can be imported. The graphical representation of the environment appears similar in quality to that of Gazebo. 2.3 Discussion of the State of the Art A successive assessment of this state of the art shows that open source solutions already offer numerous possibilities for simulating assistance and automation functions in agriculture. They are used both in research and in industry. CARLA is known for its realistic 3D environment and comprehensive sensor models, while Gazebo impresses with its versatile application possibilities in robotics. However, both tools are not specifically developed for the agricultural sector. In particular, the graphical representation and physics simulation do not meet the latest standards set by Unreal Engine 5, for example. In addition, both platforms essentially use sensor models that generate ideal data based on ground truth information from the virtual environment, which reduces the degree of realism of the sensor data that can be generated. In the context of the field of application described here, Blender can only be used for pure visualization purposes. There is potential for further development in order to better meet the requirements of agricultural simulations. The commercial simulation tools are primarily anchored in the automotive sector. dSPACE and IPG Automotive offer comprehensive simulation solutions that can also be used in an agricultural context, but there is a lack of concrete application examples in published research and industry. Both have extensive functions for modeling vehicle dynamics, sensors and scenarios according to common standards. AgriSI from Soluzioni Ingegneria, on the other hand, is specifically geared towards the simulation of agricultural vehicle functions, but offers a smaller range of functions. The graphical representation in dSPACE and IPG is qualitatively similar to that of CARLA or Unreal Engine 4, but the rendering and physics engines used remain unclear. The use of Unreal Engine 5 or comparable, latest game engines could contribute to better graphics and physics mapping. Overall, these commercial tools 2. Fachtagung TestRig - September 2024 21 Innovative Test Field Approach for Agricultural Applications offer powerful solutions, but are not specifically adapted to the development of function for agriculture. All of the virtual simulation environments mentioned deal only marginally with mobile agricultural machinery technology. Only AgriSI deals with this topic in a relevant way, but does not provide a complete simulation environment for assistance and automation functions. The frequently mentioned configuration of scenarios according to ASAM standards offers practical possibilities in the automotive environment, but is designed for off-road applications. 3. Demand-Orientated Synthesis of System Properties The definition of system properties of a VTF for mobile agricultural applications forms the starting point for the development of corresponding software. First of all, the assistance and automation functions to be developed and tested with the help of the VTF require function-related properties of VTF. In addition, there are method-related properties that reflect the product developments perspective and show how a VTF must be designed in order to effectively support the development process. A process to synthesize system properties for VTF is shown in Fig. 3. It starts with the state of the art and the current challenges in agriculture. Subsequent consideration of the functions to be developed and the needs of the product developers result in desired properties of the VTF, which fulfill the quality requirements for software from ISO/ IEC 25010 [31]. Fig. 3: Process to synthesize Software Properties based on current Challenges in Agriculture and the State of the Art 3.1 Challenges and Solution Approaches A primary challenge in agriculture is the workforce shortage, driven by demographic changes and declining interest in agricultural careers. Statistics indicate a sharp decrease in agricultural workers in Germany over the last few years [32]. Automation presents a solution, reducing the need for personnel and enhancing job appeal by replacing strenuous tasks with monitoring roles. Additionally, the increasing complexity of agriculture necessitates continuous learning and adaptable technical systems. Legal mandates, such as the EU’s Common Agricultural Policy, require extensive understanding, documentation and monitoring, often causing farmer protests [33]. Furthermore, climate change exacerbates costs by introducing new pests and diseases. Automated systems for process monitoring and documentation can alleviate these complexities. Efficiency is another critical issue. European farmers face pressure from global and low-cost foreign competitors to optimize processes and cut costs. Enhanced efficiency, through parallelization, remote control, and minimizing damage and downtime, can lower consumer food prices, aiding in mitigating European inflation [34]. 3.2 Mechatronic Functions as a Solution Approach In the field of agricultural vehicles and implements, mechatronic functions for assistance and automation support the implementation of these solutions. For example, the use of partially automated or even autonomous vehicles and combinations, as in the “Combined Powers” project, reduces the need for personnel [35]. The use of artificial intelligence as an essential part of the system software is also mentioned there. Functions for the remote control of vehicles also offer an opportunity to relieve the farmer. In addition to the driving task, manipulation tasks of the towing vehicle and attachment can also be automated. Examples include automated soil cultivation, sowing and plant protection or functions for predictive maintenance, such as monitoring the wear condition of tools on the implements [36]. The combination of autonomous driving with other automated functions for manipulation or monitoring increases the complexity. Furthermore, the data volumes generated within such systems can be used specifically for process monitoring and documentation, which leads to a better overview of work processes and supports their planning. When looking at the described assistance and automation functions, it becomes clear that all examples are mechatronic systems according to [3]. It also mentions that mechatronic systems are typically developed model-based and with the help of simulation. Here, virtual implementation, virtual commissioning and virtual testing are explicitly described as part of the development process. As mentioned above, the use of artificial intelligence for vehicle functions is also of great importance. The “ADAS/ AV Development Lifecycle” proposed in [37] illustrates the importance of data generation and simulation for the development of such intelligent functions. In order to adequately support the development of mechatronic systems with artificially intelligent components, the VTF must integrate the function to be developed into a highly realistic, virtual operating environment. There, the realistic testing and data-driven development of the described functions can take place in a safe and efficient environment. The desired system properties of VTF can be derived from this task. In order to ensure a generally unders- 22 2. Fachtagung TestRig - September 2024 Innovative Test Field Approach for Agricultural Applications tandable formulation of the properties, they are assigned in the quality requirements for software systems defined according to the product quality model of [31]. This model is shown in Fig. 4. All these requirements are discussed in the next two sections and supplemented by specific target properties of virtual test fields. Fig. 4: Product Quality Model according to [31] 3.2 Properties for achieving Functional Suitability The most comprehensive properties are those relating to the “Functional Suitability” requirement. The functions of the VTF as a tool must make it possible for people involved in the operation, such as drivers or instructors, to be included in the simulation. The VTF to be configured should enable the integration of hardware, for example in the form of XiL methods. In addition, it should be possible to integrate functions implemented in ROS into the VTF. Its structure should consist of variable modules and reflect the structure of mechatronic systems. Data exchange within the VTF should also be based on that in the real mechatronic system. The communication of the software function to be developed with external information systems is to be represented by models of the same. Alternatively, existing external information systems can be integrated via interfaces. Further functional properties of the VTF describe the modeling of the vehicle and its functions. Various sensor principles, such as LIDAR or cameras, must be modeled to simulate environment detection. These models should provide highly realistic raw data. To achieve this, it must be possible to model realization effects such as soiling, reflection, light refraction, distortion effects of the sensor or other optical phenomena. As with the solutions from the state of the art, equivalent ground truth data should be generated for annotating the raw data in order to create labeled data sets for machine learning. Sensors that are not used for environment detection should also be modelled with physical evidence. As they are of particular importance for the functions described above, the sensor technology is listed separately here. For a complete simulation, the other components of the mechatronic system must also be modeled in a physically evident way. This also includes models for mapping actuators, driving dynamics or manipulative elements. With regard to the modeling of the virtual environment, some properties can also be determined. For example, the operational environment must be reproduced three-dimensionally as models with detailed contours (geometries) and coloring (texture) in order to be able to simulate a realistic environment. This includes static elements, such as vegetation or buildings and dynamic elements, such as people or animals. Another important point, particularly for agricultural machinery technology, is the realistic representation or manipulability of the ground. Environmental conditions, such as weather and the position of the sun depending on the time of day, must also be mapped virtually. Furthermore, the geometric model of the vehicle or combination is included in the environment models. Standards, such as those in [12], should also be observed, provided they are practicable for the agricultural field of application. The state of the art also shows that the latest game engines, such as the Epic Games Unreal Engine 5, should be used for visualization and physics emulation. Furthermore, physical phenomena should not only be realizable using an engine-based physics emulation. There should be the option to design and implement physical processes manually. This allows the user to integrate customized dynamic models, such as for vehicle dynamics, into the simulation. In addition to model the functions to be developed and their environment, a VTF must provide options to generate data for the development of machine-learning systems, as many assistance and automation functions contain such components. For structured testing, the scenarios must also be flexibly configurable, as the solutions from the state of the art already show. In addition, the automation of virtual scenarios and test processes is an essential feature for the operational efficiency of the VTF and must therefore be guaranteed. For an accurate simulation of real agricultural scenarios, it must be possible to recreate routes virtually on the basis of GPS coordinates. 2. Fachtagung TestRig - September 2024 23 Innovative Test Field Approach for Agricultural Applications Fig. 5: Block Diagram of the Mechatronic System as defined in [3] 3.3 Properties to fulfill further Requirements In addition to functional suitability, there are other requirement that are particularly important for the developers using the system. As the function of the VTF is the main focus of this article, they are only briefly mentioned here. One of these requirements is “Performance Efficiency“, which is essential for the developer. The VTF simulations must perform well enough to replace real-world tests and development steps. It should operate on local powerful workstations and be designed for real-time capable simulations. The VTF must fulfill the “Compatibility” requirement, having interfaces for combined operation with existing simulation environments and meeting recognized standards. Common protocols like ISOBUS should be considered for hardware integration. To meet “Interaction Capability”, the VTF must be clear and user-friendly for development engineers, ensuring transparent, explainable and traceable processes. Documentation of development processes, including reports and support for requirements management are essential. Furthermore, Hardware integration support, such as in terms of XiL testing, must be provided. Moreover, “Reliability” is crucial for the VTF’s robust operation, ensuring reproducibility of simulations, minimizing data loss during crashes and securing restarts. The VTF must meet “Security” and “Safety” requirements, operating offline to avoid hacker threats, respecting personal rights in sensor data and generating meaningful reports during malfunctions. Finally, “Maintainability” and “Flexibility” are essential for long-term use. VTF modules should be interchangeable, flexibly configurable and support co-simulation environments, ensuring software used is freely configurable, such as with C++ programming in the CARLA Simulator. Common communication standards must be provided for integrating external information systems. 4. Approach on a Virtual Test Field Architecture In the following section, an architecture for VTF is proposed on the basis of the previously formulated properties. Suggestions for implementation are also given. As VTF are a development tool, their architecture depends both on the type of products being developed and on the functionality requirements. 4.1 Classification of the Functions to be developed The first step is therefore to look at the functions to be developed, which are assistance and automation functions for use in agricultural vehicles and machines. and can be assigned to the class of mechatronic systems as defined in [3]. This fact is important while deriving the architecture and is therefore described here. For a better understanding of the following argumentation, the structure of mechatronic systems is illustrated in Fig. 5 based on this standard. Mechatronic systems are integrated overall systems consisting of mechanical, electrical, electronic and other types of components that perform a specific function. Mechatronic systems use sensors to perceive their environment and the status of their basic system. The signals from the sensors are processed in a function information system. Data from human-machine interfaces and external information systems also flows into this system. The function information system also generates signals to control the actuators. These act on the basic system and change its status in the direction of a target status. By linking the sensors, information system, actuators and basic system, the mechatronic system represents a closed-loop system. 24 2. Fachtagung TestRig - September 2024 Innovative Test Field Approach for Agricultural Applications The connection of the components through material, energy and information flow is indicated by arrows in Fig. 5. An agricultural example of a mechatronic system is an automatic hoeing machine that is used to remove weeds in row crops of field plants. It records the field with cameras as sensors, whereupon the weeds are identified by software as part of a function information system and the electric motors as actuators are triggered. The mechanical system of the hoe (basic system) then removes the weeds. The driver of the tractor controls the process via a panel (human-machine interface). The hoeing machine also communicates with the tractor’s control units (external information systems) via bus systems. Fig. 6: Block Diagram of an Architecture for Virtual Test Fields as modular Framework 4.2 Synthesis of a modular Architecture The described structure of the mechatronic system consisting of several components forms the basis for the structural architecture of the VTF presented in this article. The function information system component of the mechatronic system forms the software product whose development process is to be supported by the VTF. In order to be able to test and develop this software component in an integrated manner, it is useful to embed it in a modular environment consisting of models of the other components of the mechatronic system. The properties of the models are based on the description of the components from [3], but must also be interpreted further. Besides the models of all components of the mechatronic system, the VTF must include further modules for control and management by the user. It is therefore clear that the architecture of the VTF must be designed as a modular framework. A proposal for a modular framework is shown in Fig. 6. It emerges from the mechatronic system by replacing its components with models, adding further modules and adapting connections. The modules of the test field architecture are described hereinafter. The environment models are a highly realistic representation of the function’s operating environment in terms of its visualization and physics. It contains all necessary elements and enables various scenarios and environmental conditions to be set for comprehensive virtual testing. Models of the used sensor technology generate realistic and format-appropriate information for the function information system. Depending on its type, level of maturity and stage of function development, this information can be designed in different ways. A further distinction is made in the sensor models according to the origin of the information to be recorded. The System Sensor Models simulate sensors for recording internal states of the Basic System. The Environment Sensor models correspond to sensors for recording the environmental states or the interaction between the system and the environment. This separation is justified on the one hand by the high relevance of environment detection for assistance and automation functions and on the other hand by differences in the implementation of both sensor classes. Both classes together form the Sensor Model module of the VTF. The function information system is embedded in the VTF as a software product to be developed. It represents the software of the mechatronic assistance and automation functions at various stages of development. Based on control signals from the function information system, the actuator models act on the basic system models. The basic system models are a virtual replica of the basic system whose state is to be influenced. In contrast to the description from [3], it is assumed that the basic system models also interact with the environment models. This is because it is specified that the basic system is the part of the function and machine that has a manipulative effect on the environment and reacts to its influence. External information systems should be able to be integrated into the VTF via a communication interface 2. Fachtagung TestRig - September 2024 25 Innovative Test Field Approach for Agricultural Applications and humans are integrated into the test field via two different interfaces. While in [3] the human only represents the user of the function, his task in the VTF is more diverse. This is because here the human also is the product developer who uses the VTF and therefore has monitoring and controlling tasks. But for testing and development purposes, he can also take the roles of users, such as that of the driver or of a pedestrian. In the role of a user, the human is integrated by a Human-Machine Interface, which can differ depending on kind of this role. As a developer, the human accesses the functions of the Management and Utilities module via a User Interface (UI). The Management and Utilities module equips the VTF with the functions of a development tool and serves as a control center for the developer. The developer can use the module to adapt the individual models and set their parameters. It is used to configure scenarios and acts as a mediator for the dynamic control of the tool during ongoing simulation processes. Furthermore, functions for automating test scenarios are essential to ensure efficient usability. The module stores recorded simulation data in a structured manner. Moreover, there are functions for monitoring virtual tests and reporting results. Depending on the function to be developed, some of the modules described above can be omitted. For example, functions that merely provide the driver with information do not require models of the actuator or basic system. There are two types of information flows in the structure described here. As the VTF is a model-based tool, the material and energy flow of the mechatronic system are part of the information flow. They are summarized in the model information flow, via which emulations and models communicate with each other. The exchange of information between the human developer and the function information system also takes place via this channel, as this is information that occurs in a similar form in the real mechatronic system. The transfer of information relevant to development and testing takes place via development information flow. This data includes data on simulation results, module configuration, monitoring and control of the simulation as well as other data. As a bidirectional channel, it connects the Management and Utilities module with all other modules. 4.3 Assignment to Software Architecture Patterns The quality of the structure described can be confirmed by its assignment to common architecture patterns for software systems. As the solutions from the state of the art are very similar to the VTF described here in terms of their function, an examination of their architectures can serve as a starting point for the classification of the VTF. Both CARLA and the Gazebo Simulator use a client-server architecture in which the server handles the rendering and the physics engine. In both cases, the clients control the environmental conditions, their elements and actors as well as all interactions. [38, 39] There is no information on the software architecture for commercial solutions from dSPACE and IPG Automotive. However, if the solutions of these companies aim to achieve a similar functionality to that of the VTF outlined here, several of their tools must be used in combination. It can be assumed that some of these tools use similar architectures to the open source solutions. As with the commercial solutions, the VTF described in this paper also works across several tools and even hardware. The various tools involved each function according to their own software architecture, which can be assigned to a specific architecture pattern. The Environment Models module can be implemented in a similar way to CARLA and Gazebo. Therefore, the client-server architecture pattern might occur. A component-based structure as described in [40] lends itself as an overarching approach for the distributed software system of the VTF. This software does not explicitly have a software architecture pattern, but a model that divides software into reusable, interchangeable and independent components, with each component having well-defined interfaces. This makes them integrable, interchangeable and interoperable. When implementing the VTF, principles of component-based and integrated software architectures software engineering should therefore be used [40]. 4.4 Initial considerations for implementation Components of mechatronic systems differ greatly from one another in their technical and physical domains. Specialized methods and software tools must therefore be used for the model-based implementation of the components in the test field. The product developers’ perspective also underlines the need to use different software for a specialized modelling of different components. This is because only an implementation with specialized software that exactly matches the components tasks is sufficiently precise and robust to be fully justifiable on the customer side. Numerical simulation software, for example, is the recognized standard for the mathematical modelling and simulation of dynamic systems. For example, they can implement the VTFs modules for actuator models, basic system models and system sensor models of the Sensor Model module. Game engines, such as Unreal Engine 5, are particularly suitable for modelling environments, as they can display highly realistic visualization and physics emulations. Environment Sensor Models can also be integrated into these game engines via programming interfaces. This description makes it clear that a different implementation justifies the division into System and Environment Sensor Models within the Sensor Model module. Furthermore, it should also be possible to simulate modules that are implemented on different hardware platforms in combination. For example, in order to implement XiL methods. This must be taken into account when defining the model information flow, among other things. The flexible exchange of modules or the modification of their implementation is another important feature to ensure applicability with different mechatronic systems. Another important aspect is the exchange of data within the test field and beyond its boundaries. Various approaches to communication are available here. Depending on the hardware used and the distribution of the module soft- 26 2. Fachtagung TestRig - September 2024 Innovative Test Field Approach for Agricultural Applications ware on this hardware, different protocols are suitable for communication. For example, the Network Device Interface (NDI) from NewTek is highly suitable for the high-performance transmission of video data [41]. The User Datagram Protocol (UDP) is also suitable for data transmission. It is already used within a VTF for from the automotive sector in [42]. The Secure Reliable Transport Protocol (SRT) can also be used to enable communication with hardware that uses ARM processors. Furthermore, there are various options for human-machine interfaces to integrate the human developer, depending on the task. These range from simple operation via keyboard, mouse and monitor to complete integration of the developer in the simulation with the aid of motion capture and head-mounted displays, as shown in [42]. The cross-tool and cross-hardware development and test environment described above basically meets the quality requirements described in the previous section. It has the system properties assigned to the requirement. If the design is specified more precisely, the requirement can be checked for compliance and a more extensive assignment to common software architecture patterns and integration approaches can be made. 5. Practical Application Examples As part of the current development, various application examples are being created that implement the modules of the VTF in practice. The focus is on the design of the virtual environment and systems for the perception of the environment are the subject of consideration. A first application example is the implementation of virtual test rides in highly realistic agricultural environments. The central module of the VTF to be developed for this purpose is the Environment Model shown in the system diagram and image of Fig. 7. Its visualization forms the basis for testing functions with Environment Sensors. The functions of the Management and Utilities module also play a role. The starting point is the construction of an agricultural scenario in the virtual environment of Epic Games Unreal Engine 5. With the help of the game engine editor, a large map of the environment with a varied topology can be easily created. Rivers and roads are displayed, while the horizon is delimited by modeled mountains. The level of detail of this environment map is kept low for performance reasons. For virtual test rides, a very detailed operational area is integrated into the large, low-detail map of the surroundings. At its center is a field with a varied contoured and textured surface. Part of the field is planted with crops in the form of grain. Other parts are equipped with grass stubble and represent a fallow area. With regard to the environmental conditions, the configuration of the times of day has been implemented. Other weather conditions, such as the degree of cloud cover, fog or precipitation, can be integrated. The field is surrounded by buildings, trees, woodland and overgrown farmland, which visually separate it from the surrounding map and further limit the horizon. A model of a tractor is also placed on the field. The underlying programming enables it to automatically follow trajectories of agricultural maneuvers based on GPS data. With this setup consisting of a virtual environment and an automated vehicle model, virtual test rides can be carried out. After equipping the vehicle model with Environment Sensor Models, sensor data for testing environmental perception functions can be generated in this scenario. Fig. 7: Virtual Test Ride in a highly realistic Agricultural Environment with Models of Plow and Tractor [43, 44] The Environment Model presented above, an Environment Sensor Model and Management and Utilities functions play a role in the next two examples, what is indicated by the system diagram in Fig. 8. Fig. 8: Modules used for sensor-based Applications Since the virtual environments of the VTF offer highquality graphical representation, they are well-suited for simulating camera-based or image-based applications. An example of such an application in agricultural machinery is the camera-based monitoring of tools on implements for soil cultivation [36]. Another practical application are positioning systems for field hoes used for weed 2. Fachtagung TestRig - September 2024 27 Innovative Test Field Approach for Agricultural Applications removal which differentiate between crops and weeds based on image recognition systems [45]. Such systems are of fundamental importance, especially for the operation of implements in conjunction with future autonomous tractors, where no driver will oversee or control the implements directly. The algorithms used in these systems for image evaluation are developed and tested using image data. This image data can be generated efficiently, reproducibly and following clear environmental scenarios through virtual drives in the VTF. However, the image data must closely correspond to the representations of the camera used in the real system, which is why the VTF camera model must be configured based on the calibration data of the real camera. The synthetic image data generated with such a model accurately replicate real recordings in terms of image distortion and field of view. Fig. 9 shows the image rendered by the game engine and the realistic replication distorted based on calibration data. Fig. 9: Ideal and realistically distorted synthetic Images A further example is the generation of test and training data sets for image-based machine learning object recognition algorithms in virtual environments. For machine learning, these data sets must contain raw sensor data in the form of images of the objects to be recognized. Furthermore, this raw image data must be annotated by labeling the positions and classes of the objects in the image recordings with so-called bounding boxes and assigning the objects to an object class. This is done in the VTF by querying the class and position of the object models in the camera’s field of view and storing this ground truth information together with the raw images. Fig. 10 shows the image data annotation process in a simplified form. While in many applications the annotation of image data is still done manually, the implementation in the VTF enables the complete automation of this process. The data generator described here can be operated in parallel to the simulation in order to test object detectors in real time and in flexible scenarios. Furthermore, the synthetically generated data sets from annotated images can be used to develop algorithms for object detection. Fig. 10: Data Generation Process for image-based Machine Learning Object Recognition Algorithms A test of this approach is conducted for demonstration purposes. It results in a large number of annotated images of a virtual horse model, as it is shown Fig. 10. This synthetic image data forms 90% of a hybrid data set used to train a machine learning object detector based on the Yolo algorithm [46]. Fig. 11 shows the evaluation of a real image using the object detector designed in this way. It is evident that the object detector, which is largely trained with synthetic data, is capable of precisely evaluating real images. Fig. 11: Result of Object Detection in a real Image from [47] with a Detector trained with Hybrid Data 28 2. Fachtagung TestRig - September 2024 Innovative Test Field Approach for Agricultural Applications An outlook on further application examples is given in the following section. However, the examples shown are already sufficient to emphasize the functionality and usefulness of the VTF for testing and product development. 6. Conclusion This study shows that VTF are a promising solution for development of sophisticated agricultural applications in order to cope with the increasing demand on automation and therefore on suitable testing methods. The presented approach enables reproducible test scenarios, temporal independence and flexible virtual testing, which can still be further adapted to the specific requirements of agriculture. Based on the current state of the art, it was possible to derive an architecture that meets standardized quality requirement. This means that the described VTF can not only be used scientifically, but is also justifiable as a robust solution within industrial development processes. In general, the future prospects for VTF in mobile agricultural machinery development and commercial vehicle technology are very promising. Advancing digitalization will further increase the demand for such tools. Especially new technologies, such as machine learning will expand both the demand and application possibilities of VTF. This will support the development of efficient and sustainable vehicle and machine systems and contribute to overcoming future challenges in agriculture. Further developments can also be expected in the specific case of this VTF. For instance, the presented examples will be developed further in the direction of industrially usable tools in the future. Further functions will also be added and integrated into the VTF. One existing concept, for example, is the integration of dedicated, prototypical hardware for environment perception functions into the VTF. This requires the creation of suitable, high-performance interfaces and an evaluation of usability. This step is intended to create the basis for the implementation of processor-in-the-loop methods in the VTF. Further investigations, evaluations and technical developments of the modules are also planned as well as a transfer of the principle to other commercial vehicle applications. References [1] O. Burkacky, J. Deichmann, and J. P. Stein, Outlook on the automotive software and electronics market through 2030. [Online]. Available: https: / / www.mckinsey.com/ industries/ automotive-and-assembly/ our-insights/ mapping-the-automotive-software-and-electronics-landscape-through-2030 (accessed: May 19 2024). [2] C. Steinbrink et al., “Simulation-Based Validation of Smart Grids - Status Quo and Future Research Trends,” in Lecture Notes in Computer Science, Industrial Applications of Holonic and Multi-Agent Systems, V. Mařík, W. Wahlster, T. Strasser, and P. Kadera, Eds., Cham: Springer International Publishing, 2017, pp. 171-185. [3] Development of mechatronic and cyber-physical systems, VDI/ VDE 2206, VDI/ VDE Society Measurement and Automation, Düsseldorf, 00.Nov. 2021. [Online]. Available: www.vdi.de/ richtlinien/ programme-zu-vdi-richtlinien/ vdi-2206 [4] M. Placek, Autonomous Vehicle Simulation Solutions Market - A Global Market and Regional Analysis. [Online]. Available: https: / / bisresearch.com / industry-report/ autonomous-vehicle-simulationsolution-market.html? utm_source=Statista (accessed: May 19 2024). [5] L. Goedde, J. Katz, A. Ménard, and J. Revellat, Agriculture’s connected future: How technology can yield new growth. [Online]. Available: https: / / www.mckinsey.com/ industries/ agriculture/ our-insights/ agricultures-connected-future-how-technology-can-yield-new-growth (accessed: May 19 2024). [6] M. Cutini et al., “A Co-Simulation Virtual Reality Machinery Simulator for Advanced Precision Agriculture Applications,” Agriculture, vol. 13, no. 8, p. 1603, 2023, doi: 10.3390/ agriculture13081603. [7] Institute of Electrical and Electronics Engineers (IEEE), Author Guidelines for Integrative Literature Reviews. [Online]. Available: https: / / procomm. ieee.org/ wp-content/ uploads/ 2019/ 05/ Author-Guidelines-for-Integrative-Literature-Reviews.pdf (accessed: Jun. 25 2024). [8] CARLA Team, CARLA: Open-source simulator for autonomous driving research. [Online]. Available: https: / / carla.org/ / (accessed: Sep. 9 2024). [9] I. Epic Games, Unreal Engine: The most powerful real-time 3D creation tool. [Online]. Available: https: / / www.unrealengine.com/ de (accessed: Jun. 9 2024). [10] CARLA Team, Let’s go off-road! : AVL has integrated CARLA into their simulation toolchain for autonomous driving use cases in unpaved terrain. [Online]. Available: https: / / carla.org/ 2023/ 04/ 21/ avl-off-road-simulation/ (accessed: Jun. 9 2024). [11] T. Dellmann and K. Berns, “Toward a Realistic Simulation for Agricultural Robots,” in Smart Innovation, Systems and Technologies, Agriculture Digitalization and Organic Production, A. Ronzhin, K. Berns, and A. Kostyaev, Eds., Singapore: Springer Nature Singapore, 2022, pp. 3-13. [12] ASAM e.V., ASAM - Standardization for Automotive Development. [Online]. Available: https: / / www. asam.net/ about-asam/ (accessed: Jun. 9 2024). [13] Open Source Robotics Foundation, Inc, GAZEBO. [Online]. Available: https: / / gazebosim.org/ home (accessed: Jun. 9 2024). [14] Open Source Robotics Foundation, Inc, ROS - Robot Operating System. [Online]. Available: https: / / www.ros.org/ (accessed: Jun. 13 2024). [15] I. Clearpath Robotics, Additional Simulation Worlds. [Online]. Available: https: / / www.clearpathrobotics.com/ assets/ guides/ kinetic/ husky/ additional_sim_worlds.html (accessed: Jun. 9 2024). 2. Fachtagung TestRig - September 2024 29 Innovative Test Field Approach for Agricultural Applications [16] PickNik Inc., Robot Operating System (ROS). [Online]. Available: https: / / picknik.ai/ ros/ (accessed: Jun. 9 2024). [17] A. Ivanovic, M. Polic, J. Tabak, and M. Orsag, “Render-in-the-loop Aerial Robotics Simulator: Case Study on Yield Estimation in Indoor Agriculture,” in 2022 International Conference on Unmanned Aircraft Systems (ICUAS), Dubrovnik, Croatia, Jun. 2022 - Jun. 2022, pp. 787-793. [18] R. J. H. Kemper, C. Gonzalez, and S. R. P. Gardini, “Autonomous Navigation of a Four-Wheeled Robot in a Simulated Blueberry Farm Environment,” in 2022 IEEE ANDESCON, Barranquilla, Colombia, Nov. 2022 - Nov. 2022, pp. 1-6. [19] R. S. Raut, G. B. Mahanta, A. Rout, S. K. Pattanayak, and B. B. Biswal, “Design and Simulation of a Multi-purpose Adjustable Modular Robot for Precision Agriculture,” in Smart Innovation, Systems and Technologies, Design in the Era of Industry 4.0, Volume 2, A. Chakrabarti and V. Singh, Eds., Singapore: Springer Nature Singapore, 2023, pp. 915-927. [20] Blender Foundation, Blender: The Free and Open Source 3D Creation Suite. [Online]. Available: https: / / www.blender.org/ about/ (accessed: Jun. 9 2024). [21] J. Liu, Y. Zhao, X. Xiao, R. Meng, and J. Liu, “CAS: Crop Aerial Sensing Simulation in Smart Farming,” in 2023 20th Annual IEEE International Conference on Sensing, Communication, and Networking (SE- CON), Madrid, Spain, Sep. 2023 - Sep. 2023, pp. 48-53. [22] dSPACE GmbH, dSPACE: Partner in Simulation and Validation. [Online]. Available: https: / / www.dspace.com/ en/ inc/ home/ company.cfm (accessed: Jun. 9 2024). [23] dSPACE GmbH, Aurelion. [Online]. Available: https: / / www.dspace.com/ de/ gmb/ home/ products/ sw/ experimentandvisualization/ aurelion_sensor-realistic_sim.cfm (accessed: Jun. 19 2024). [24] dSPACE GmbH, VEOS. [Online]. Available: https: / / www.dspace.com/ de/ gmb/ home/ products/ sw/ simulation_software/ veos.cfm (accessed: Jun. 19 2024). [25] dSPACE GmbH, Autonomous Farming: Test Autonomous Farm Machines in All Seasons. [Online]. Available: https: / / www.dspace.com/ en/ pub/ home/ products/ sw/ experimentandvisualization/ aurelion_ sensor-realistic_sim/ autonomous-farming.cfm (accessed: Jun. 20 2024). [26] dSPACE GmbH, ModelDesk. [Online]. Available: https: / / www.dspace.com/ de/ gmb/ home/ products/ sw/ automotive_simulation_models/ produkte_asm/ modeldesk.cfm (accessed: Jun. 19 2024). [27] IPG Automotive GmbH, Solutions for Virtual Test Driving. [Online]. Available: https: / / www.ipg-automotive.com/ de/ (accessed: Jun. 20 2024). [28] S. Bilgic Istoc, Automatisierte Nutzfahrzeuge: Herausforderungen in der Simulation. [Online]. Available: https: / / www.hanser-automotive.de/ a/ fachartikel/ herausforderungen-in-der-simulation-3047932 (accessed: Jun. 20 2024). [29] Soluzioni Ingegneria Srl, AgriSI: -Smart Technologies for Agriculture can now be tested in Simulation. [Online]. Available: https: / / si-ita.it/ products/ agrisi/ (accessed: Jun. 9 2024). [30] Tractors and machinery for agriculture and forestry. Serial control and communications data network. General standard for mobile data communication, ISO 11783, ISO, Geneva, Jan. 2018. [31] Systems and software engineering: Systems and software Quality Requirements and Evaluation (SQuaRE), ISO/ IEC 25010, ISO / IEC, Geneva, 00.Nov. 2023. [32] German Federal Statistical Office, Pressemitteilung Nr. 021. Wiesbaden, 2024. Accessed: Jun. 6 2024. [Online]. Available: https: / / www.destatis.de/ DE/ Presse/ Pressemitteilungen/ 2024/ 01/ PD24_021_ 41.html [33] A. Koper and P. Blenkinsop, EU eyes easing green rules on farmers to defuse protests. [Online]. Available: https: / / www.reuters.com/ world/ europe/ eu-planease-farmers-fallow-land-requirement-stalls-detail- 2024-03-15/ (accessed: Jun. 6 2024). [34] FAO, World Food and Agriculture - Statistical Yearbook 2023: FAO, 2023. Accessed: Jun. 6 2024. [Online]. Available: https: / / openknowledge.fao.org / items/ 5c272dc7-e1b8-486a-b323-6babb174eee0 [35] S. Morris, M. Steinwendner, and M. Ehses, Driverless in the field: Agriculture companies KRONE and LEMKEN rely on software from Apex.AI for autonomous farming machines. Palo Alto, 2023. Accessed: Jun. 6 2024. [Online]. Available: https: / / combined-powers.com/ en/ downloads-and-pr/ [36] LEMKEN GmbH & Co. KG, Camera for monitoring cultivator share wear. Alpen, 2023. Accessed: Jun. 6 2024. [Online]. Available: https: / / lemken.com / en-en/ news/ agriculture-news/ detail/ camera-for-monitoring-cultivator-share-wear? utm_campaign=iqtool-monitoring&utm_content=teaser&utm_medium=email&utm_source=newsletter [37] Microsoft Corporation, Reference architecture for autonomous vehicle operations (AVOps). [Online]. Available: https: / / learn.microsoft.com/ en-us/ industry/ mobility/ architecture/ ra-mobility-avops (accessed: Jun. 13 2024). [38] CARLA Team, Foundations. [Online]. Available: https: / / carla.readthedocs.io/ en/ latest/ foundations/ (accessed: Jun. 16 2024). [39] Open Source Robotics Foundation, Inc, Gazebo Architecture. [Online]. Available: https: / / classic.gazebosim.org/ tutorials? tut=architecture&cat=get_started (accessed: Jun. 16 2024). [40] I. Sommerville, Software Engineering, 10th ed. Hallbergmoos: Pearson, 2018. [41] NDI, There is no video without connection. [Online]. Available: https: / / ndi.video/ tech/ (accessed: Jun. 16 2024). 30 2. Fachtagung TestRig - September 2024 Innovative Test Field Approach for Agricultural Applications [42] R. Degen et al., Integration of Driving Physical Properties into the Development of a Virtual Test Field for Highly Automated Vehicle Systems. [43] J. Stein, John Deere 6195 Series Tractor, Detailed With SCV Ports, Drawbar Hitch, & 3PT. [Online]. Available: https: / / grabcad.com/ library/ john-deere- 6195-series-tractor-detailed-with-scv-ports-drawbar-hitch-3pt-1 (accessed: Jul. 11 2024). [44] cjkef hvkflkvçefwvjl, CULTIVATOR. [Online]. Available: https: / / grabcad.com/ library/ cultivator-17 (accessed: Jul. 11 2024). [45] Garford Farm Machinery Ltd, Robocrop Guided Hoes. [Online]. Available: https: / / garford.com / products/ robocrop-guided-hoes/ (accessed: Jul. 14 2024). [46] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, Jun. 2016 - Jun. 2016, pp. 779-788. [47] I. Roboflow, Horse Dataset. [Online]. Available: https: / / universe.roboflow.com/ z-firewood-house/ horse-v82oc/ dataset/ 1 (accessed: Jul. 11 2024).