Results of the AGRARSENSE and AIFor projects at Lapland University of Applied Sciences

Editor:

  • Saloniemi Anne, MNR (forestry), Senior Specialist, Future Bioeconomy Expertise Group, Lapland University of Applied Sciences

Authors of the articles:

  • Avula Ramana Reddy, PhD, Senior Researcher, Dependable and Autonomous Systems Unit, RISE Research Institutes of Sweden
  • Korhonen Minna, PhD, Senior Lecturer, Electrical Engineering Expertise Group, Lapland University of Applied Sciences
  • Narkilahti Aleksi, B. Eng, Specialist, Digital Solutions Expertise Group, Lapland University of Applied Sciences
  • Saloniemi Anne, MNR (forestry), Senior Specialist, Future Bioeconomy Expertise Group, Lapland University of Applied Sciences
  • Unga Jussi, B. Eng, Specialist, Digital Solutions Expertise Group, Lapland University of Applied Sciences

Preface:

  • Saloniemi Anne, MNR (forestry), Senior Specialist, Future Bioeconomy Expertise Group, Lapland University of Applied Sciences

Summary

This compliation presents the main results of the Smart, digitalized components and systems for data-based Agriculture and Forestry (AGRARSENSE) and AI training of automated forest machine (AIFor) projects, in which Lapland University of Applied Sciences (Lapland UAS) participated to advance digital solutions for forestry. The articles describe how high‑fidelity simulation, digital twins, and synthetic data can support the development, training, and validation of AI‑based perception and autonomous functions for forestry machines—improving safety, efficiency, and cost‑effective R&D in complex forest environments.

This publication is intended for RDI professionals, researchers, educators, technology developers, and industry stakeholders interested in forestry automation, digital twins, and AI validation. By compiling methods, lessons learned, and concrete outputs, the publication supports the uptake of results in research, education, and future development projects beyond AGRARSENSE and AIFor. Most authors are specialists and lecturers from Lapland UAS, with additional contributions from partner organizations.

The Lapland UAS authors represent the core team. Aleksi Narkilahti and Jussi Unga led the practical development of the Unreal Engine 5–based simulator, including sensor models, data pipelines, and virtual testing workflows. Minna Korhonen contributed expertise in point‑cloud and remote‑sensing data and its use in building and analyzing forest digital twins. Anne Saloniemi connected the technical work to forestry needs, project management, and dissemination. Ramana Reddy Avula (RISE) was invited to provide an independent validation and safety‑assessment perspective, ensuring the publication addresses not only how the simulator was built, but also how its outputs can be evaluated for credibility and used in safety‑critical use cases.

The AGRARSENSE project involved 51 partners from 14 European countries and consisted of seven use cases, including the Forestry use case in which Lapland UAS participated with several partners. This use case was led by Komatsu Forest (Sweden). AGRARSENSE was funded by the EU Chips Joint Undertaking, and AIFor was funded by Business Finland. Both projects were implemented from 1 January 2023 to 28 February 2026.

Preface

Anne Saloniemi

In recent years, the forestry sector has been expected to deliver higher efficiency while also addressing climate‑related challenges. There is a need for methods that are efficient and economically viable while also taking into account the requirements set by climate change. Current practices must be further developed to ensure the competitiveness of the Finnish and European forest sector in the global market.

To meet today’s requirements in forest machine development, the sector needs both new autonomous machine concepts and high-quality virtual environments for training and testing the AI that will control them. Yet existing simulation approaches are still too limited: current virtual forests and digital twins often lack the accuracy and level of detail needed to generate training data that transfers reliably to real operations. This is one reason why fully autonomous forest machines are not yet operating in real forests since forest environments are highly variable and difficult to sense and predict. Achieving robust autonomy therefore requires realistic digital twins and reliable sensor data both before operations (e.g., high-resolution aerial data collected with drones) and during operations using sensors mounted on the machine.

In earlier projects, automated workflows for generating virtual forests have been developed for purposes such as landscape visualization and forest planning, typically by combining existing geographic information and forest inventory data (FrostBit Software Lab 2026). However, these approaches do not yet provide the level of realism and detail required for a forest-machine simulator used in AI training. For autonomy-related use cases, the digital twin must represent the environment with higher fidelity, particularly the terrain and ground surface, which strongly affect drivability and sensing. Reaching this accuracy often requires additional, higher-resolution inputs (for example, more precise elevation and terrain data) on top of conventional GIS and forest datasets.

Digital Solutions for Forestry

New AI training methods were planned to be developed in the Smart, digitalized components and systems for data-based Agriculture and Forestry (AGRARSENSE) and AI training of automated forest machine (AIFor) projects (1, January 2023–28, February 2026). The AGRARSENSE project was based on the seven use cases shown in Figure 1: Greenhouses (UC1), Vertical Farming (UC2), Precision Viticulture (UC3), Agri robotics (UC4), Forestry (UC5), Optimal soils & fertilizers (UC6), and Water (UC7). Lapland UAS participated in the implementation of use case 5 (Forestry) with two expertise groups: Future Bioeconomy and Digital Solutions.

The Forestry use case was led by Komatsu Forest (Sweden). AGRARSENSE and AIFor addressed these challenges by improving the quality and realism of simulation and synthetic data used for training and validation, including realistic sensor responses and robustness to harsh weather.

Figure 1. Seven use cases of the AGRARSENSE project.

The AGRARSENSE project was an EU‑funded consortium under the Chips Joint Undertaking that developed advanced microelectronics, photonics, sensor technologies, and data‑management solutions for agriculture and forestry. The project was implemented in cooperation between industry partners and research institutions through international and national networks, according to needs identified by industry.

AIFor was a Business Finland–funded R&D project implemented by Lapland University of Applied Sciences (Lapland UAS). The main goal of the project was to develop new methods and a high-fidelity virtual forest environment (a digital twin) that can be used to train AI for autonomous forest machines — especially to enable reliable autonomous navigation and operation in complex, unpredictable forest conditions. The work combined detailed 3D modelling (including trees, ground vegetation, fauna and obstacles), development of a simulator platform and data pipelines, and modelling of forest machine interaction and relevant sensors (e.g., RGB cameras, depth/segmentation, LiDAR/radar) with data driven noise models to capture effects of harsh conditions such as rain and snowfall. A central element was a real site demonstration in which the digital twin and simulated outputs were compared against a real logging site to evaluate authenticity and applicability. In addition to the simulator and digital twin, the project produced a preliminary standard description for automated virtual forest generation, outlining data needs and accuracy requirements to support more efficient creation of site-specific digital twins without extensive on-site data collection, and thereby improving the scalability of simulation-based AI training and validation for forestry use cases.

At Lapland UAS, much of the practical development work was carried out within the FrostBit Software Lab, which provided the implementation environment for building the Unreal Engine 5–based forestry simulator, creating high-fidelity virtual forest content, and integrating the data pipelines and virtual sensors used in the project demonstrations. The lab’s multidisciplinary team combined game-engine development with forestry and remote-sensing expertise, enabling rapid prototyping and close collaboration with project partners.

The simulator also includes an adjustable weather and lighting system (e.g., rain, snowfall, fog, cloud cover, wind, snow depth, and time of day), with sun position calculated from the digital twin’s real-world latitude and longitude to better match field conditions during scenario design and validation. (Narkilahti 2023)

Partners benefit from Lapland UAS by gaining freely shareable, high‑quality 3D forestry assets and simulation solutions that support realistic testing, training, and validation of digital and autonomous systems (Valkama 2024).

This collection of articles is based on the main results of the AGRARSENSE and AIFor projects. The articles are written by specialists and lecturers who participated in the projects, together with one external contributor from RISE. The articles follow the AIFor project structure:

  • Developing a simulator platform for autonomous driving & AI research for forestry (WP2 Virtual Environment Development for AI training)
  • Point cloud data in simulation (WP3 Drone data use in Digital twin creation and digital twin accuracy evaluation)
  • From Digital Twin to AI-based perception system validation (WP4 Demonstration)

The publication is intended for RDI professionals, researchers, educators, technology developers, and industry stakeholders working at the intersection of forestry, automation, and digital twins, and it can be used both as a practical overview of the developed simulator and datasets and as a reference for planning and validating future simulation‑based AI development in forest environments.

Sources

AGRARSENSE, 2026. Referenced 5.5.2026 https://www.agrarsense.eu/

FrostBit Software Lab 2026. Virtual Forest 2.0. Referenced 5.5.2026. https://www.frostbit.fi/en/portfolio/virtual-forest-2-0/#:~:text=Virtual%20Forest%202.0%20is%20a%20project%20that,Rocks%20and%20rock%20lands%20*%20Felling%20areas

Narkilahti, A. 2024. Developing a simulator platform to advance autonomous systems in forestry usage. Referenced 5.5.2026. https://lapinamk.fi/blogiartikkeli/developing-a-simulator-platform-to-advance-autonomous-systems-in-forestry-usage/

Valkama, S. 2024. Making realistic 3D trees. Referenced 4.5.2025. https://lapinamk.fi/blogiartikkeli/making-realistic-3d-trees/

Developing a simulator platform for autonomous driving & AI research for forestry

Aleksi Narkilahti, Jussi Unga & Ramana Reddy Avula

Introduction

As part of the forestry use case within the Smart, digitalized components and systems for data-based Agriculture and Forestry (AGRARSENSE) and AI training of automated forest machine (AIFor) projects, the FrostBit Software Lab team at Lapland University of Applied Sciences (Lapland UAS) developed a simulator platform to support the development, training, and validation of autonomous driving technologies and AI applications for forestry. The platform was designed with two primary purposes: real-time simulation and synthetic sensor data generation. This work was carried out in close collaboration with multiple partners across several EU countries, each contributing to different aspects of the forestry use case.

The overall objective of the forestry use case was to enhance operational efficiency and safety in forest environments. At present, fully autonomous forestry machines are not yet in operation, largely due to regulatory constraints and the unpredictable and complex nature of forest environments. These challenges are comparable to those faced in urban autonomous driving, where systems must handle dynamic and uncertain conditions. The forestry use case aimed to address these challenges by improving the availability of high-quality AI training data, reducing development costs for autonomous systems, and advancing automation and safety in forestry operations.

The simulation-related work in the AGRARSENSE project was organized into two main work packages: Virtual Environment Development for AI Training (WP2) and Demonstration (WP4). WP2 focused on developing the simulator platform and its core components, while WP4 addressed validation, finalization and refinement of the simulator, and real-world demonstration activities.

This article focuses on the simulator development work carried out in WP2, including virtual environment development, modelling, sensor simulation, and forest machine interaction.

Why a custom simulator was developed

In previous simulation-related projects, the FrostBit Software Lab team made use of existing platforms such as the CARLA Simulator, which is an open-source simulator for autonomous driving research (CARLA Simulator 2026). For the AGRARSENSE project, however, a decision was made to develop a simulator platform largely from the ground up for several key reasons.

First, at the start of the project, CARLA was still based on Unreal Engine 4, while Unreal Engine 5 was required to reach the target level of visual fidelity and to maintain real-time performance. Unreal Engine 5 (Unreal Engine 5, 2026) introduced major improvements such as Nanite virtualized geometry (Epic Games 2026) for handling highly detailed assets efficiently, and Lumen global illumination for more realistic real-time lighting. These features were especially important for rendering dense, complex forest environments in a visually convincing, yet performant way.

Second, CARLA is primarily designed for urban city driving scenarios and includes many components that are unnecessary for forestry use. At the same time, it does not provide any forest environments, meaning that maps would have had to be created from scratch while leaving most of CARLA’s built-in systems unused. As a result, a custom simulator platform was developed, while leveraging the knowledge and experience gained from working with and extending the CARLA Simulator.

Developing the simulator

Before development started, the key features required for the simulator were defined. The requirements were extensive and covered several areas, such as:

  • Getting data in and out of the simulator
  • Efficient sensor simulation
  • Drivable forwarder and harvester vehicles
  • Easy-to-use interface tools for quick setup and testing
  • High‑fidelity forest environment(s) with high‑quality assets
  • Adjustable weather system

To move data in and out of the simulator, Robot Operating System (ROS) was used for communication. ROS is an open-source robotics middleware and message-passing framework (ROS 2026). The open-source ROS Integration plugin for Unreal Engine was used to communicate with the ROS framework (ROS 2026). ROS provides a broad set of data types that can be sent as messages, and support for multiple programming languages allows simulator users to work in the language of their choice. To further broaden the compatibility of the simulator, ROS 1 bridge can be used to relay messages to ROS 2. ROS 2 is a newer version of the ROS framework, but many users are still using ROS 1.

For sensor simulation, close collaboration with project partners was used to determine which sensors were most relevant for forestry use cases. The initial set included RGB, Depth, Semantic Segmentation cameras, LiDAR, Radar, as well as several virtual sensors such as transform, collision, and overlap sensors. Later in the project, this set was extended to include a Dynamic Vision Sensor, a basic thermal camera, and an Instance Segmentation camera to better support the creation of synthetic datasets. In developing these sensors, earlier experience with the CARLA Simulator was used, while most of the sensors were redesigned and simplified to make them easier to maintain and, in some cases, more performant. For example, in a previous project, CARLA’s LiDAR sensor had been improved to handle multiple LiDARs efficiently. These improvements were carried over and refined further in this project.

Both the forwarder and harvester were built from scratch as high-detail models (Figure 1). They can be driven manually or controlled via ROS. Each vehicle’s boom arm is fully controllable; the forwarder’s boom can lift and handle logs. For this project, the harvester’s sawing function is not included. Vehicle dynamics are physics-based and powered by Unreal Engine 5’s Chaos physics.

Figure 1. Forwarder and Harvester vehicles in the simulation environment

The simulator features an adjustable weather system covering rain and snowfall, cloud cover, fog, temperature, wind speed and direction, snow depth, and the month and time of day (Figure 2). The sun moves accurately for the selected date and time, calculated from the digital twin’s real-world latitude and longitude, so lighting matches actual conditions. This enables realistic scenario design from low light to bright daylight across winter, summer, and autumn.

Figure 2. Simulator weather menu.

Hirvas site visit

At the start of the project, the final demonstration site had not yet been decided. To move quickly with the modelling process and simulator development, an area in Hirvas near Rovaniemi was selected as the first digital twin map. Permission was received from the Finnish forest administration to use the site, which enabled modelling, simulator development, and map work to begin without delay.

Together with a 3D modeler and 2D artist, the site was visited to capture photos and videos of the area, along with close-ups of key objects. These materials provided essential references for creating high-quality 3D models and textures. A bit later, surveying experts together with students from Lapland UAS flew a camera-equipped drone over the area and generated a point cloud for use in the project.

Hirvas map creation

The development of the Hirvas area started with free open-source data from the National Land Survey of Finland website, where the heightmap and the canopy height model (CHM) for the tree coverage were obtained. At the start, the landscape was generated using the heightmap, and the scale and roads were cross-referenced with point cloud data from the actual location (Figure 3).

Figure 3. Hirvas landscape heightmap with road placement.

Roads and junctions were smoothed using landscape splines that adjust the landscape height to the road level. These splines are also used in the tree spawning logic as “ignore” areas, where no trees are needed. Road accuracy and placement were confirmed using road data from OpenStreetMap and the point cloud developed from images of the actual location. The Blender GIS plugin allowed actual OpenStreetMap roads from the area to be imported and exported as a mesh to Unreal Engine, where road placement could be compared.

The Finnish Forest Centre provides national forest datasets that could be used to obtain the CHM data of the Hirvas area. The CHM data was then opened in QGIS and clipped to the exact size using the previously acquired heightmap image as a clipping mask (Figure 4).

Figure 4. Hirvas heightmap on the left, and CHM data on the right.

With the correctly sized CHM data, a color range representing the tree-height distribution in the Hirvas area could be generated. In QGIS, the minimum and maximum tree heights were determined, defining the color scale from dark to bright red. Using the maximum tree height of 23 meters, QGIS converted the grayscale CHM image into the final colorized CHM map (Figure 5). This map could then be imported into Unreal Engine to place trees accurately based on real-world height data.

Figure 5. Final form of Hirvas CHM data

In Unreal Engine, the colorized CHM data (Figure 5) was used to compare each pixel’s color with its corresponding landscape coordinate and place appropriately sized trees, including an approximated species distribution. For each pixel, its color value was read, matched to the predefined color range, and used to select the closest height-appropriate 3D tree model. The selected model was then scaled up or down to introduce variation and produce a more realistic result.

Since the roads and swamp locations are baked into the CHM data without any trees, only minimal manual cleaning was required after tree generation.

Other assets such as rocks, tree stumps, sticks, pinecones, grass, and twigs were generated randomly, with much higher densities to fill out the landscape and create a more realistic appearance (Figure 6).

Figure 6. Hirvas area before and after tree generation.

Simulator sensor validation

A crucial part of the development process is to ensure that the simulator’s virtual sensors behave in a realistic and predictable manner, enabling their use for AI training, testing, and safety-related research. To achieve this, RISE performed dedicated sensor validation activities focusing mainly on the RGB camera model, as vision-based perception is central for forestry-related automation tasks such as obstacle detection, human detection, terrain interpretation, and machine guidance. Following a methodology adapted from established simulation-validation frameworks such as New Assessment/Test Method (NATM), RISE conducted controlled laboratory experiments to assess how closely the simulated camera reproduces real sensor behavior.

The validation began with standalone camera model testing using a controlled laboratory setup. A Luxonis OAK-D Pro W RGB camera was used as physical reference, and a matching virtual camera was configured in the simulator by aligning resolution, field-of-view, and exposure characteristics. A Macbeth Color Checker chart was illuminated with calibrated LED soft lights whose intensities and color temperatures (3500–5000 K) were measured using a spectrometer. The same lighting parameters were then replicated in the simulator as shown in Figure 7 to ensure comparable imaging conditions. The color accuracy of the real and simulated cameras was quantified using the CIEDE2000 ΔE metric. The virtual camera consistently achieved ΔE values between 2.56 and 3.36, indicating small perceptual differences, while the real camera exhibited larger variation due to the spectral characteristics of the LED lights.

These laboratory results demonstrate that the camera model provides a reliable basis for perception research and synthetic data generation. Scenario-based validation using real-world reference data is presented in WP4, where the camera model is evaluated within a digital twin and used for testing AI-based safety-critical functionalities.

Figure 7. Overview of the camera validation setup in the simulator.

Summary

Overall, the simulation development in the AGRARSENSE and AIFor projects progressed successfully, especially considering the challenge of building the simulator largely from the ground up. Throughout the project, collaboration with multiple partners supported simulator testing from an early stage and provided valuable feedback on improvements, required features, and bug identification. This collaboration also included RISE’s work on validating the simulator’s RGB camera model, which provided important insights into how well the virtual sensor replicates real-world behavior and where the environment could be further refined. These combined efforts were crucial in supporting the development and refinement of the simulator, and positive responses were received from the partners who used it in their work. The simulator is expected to support further use and development in the future.

Sources

AGRARSENSE 2026. Referenced 23.3.2026. https://www.agrarsense.eu/

CARLA Simulator 2026. Referenced 19.5.2026. https://carla.org/

Epic Games 2026. Nanite Virtualized Geometry Overview. Referenced 15.5.2026. https://dev.epicgames.com/documentation/unreal-engine/nanite-virtualized-geometry-in-unreal-engine

ROS 2026. Robot Operating System. Referenced 12.5.2026. https://www.ros.org/

ROSIntegration 2026. Referenced 16.4.2026. https://github.com/code-iai/ROSIntegration

Unreal Engine 5 2026. Referenced 15.5.2026. https://www.unrealengine.com/

Point cloud data in simulation

Minna Korhonen

Introduction

Modern forestry relies on advanced machinery, and developing autonomous forest machines is a growing area of robotics research. Autonomous harvesters and forwarders, supported by drones for environmental monitoring, can improve productivity and reduce soil damage. Safe operation requires machines to identify trees, humans, and animals, as well as navigate terrain with slopes, ditches, and other obstacles. Existing maps often lack precise details, so additional data sources are needed. (Visser & Obi, 2020)

Realistic simulation environments, such as virtual forests and digital twins, support the development of autonomous systems and are also used in education and forest management (Murtiyoso et al., 2023). Finland’s well-documented forests provide valuable resources, including aerial imagery and laser-scanned point clouds from the National Land Survey (Maanmittauslaitos, 2026). For local harvesting plans, more detailed models can be created using photogrammetry or LiDAR scanning (Murtiyoso et al., 2023).

A key component is the three-dimensional point cloud, which represents at least the spatial coordinates of points. A cost-effective method for generating point clouds is photogrammetry using overlapping drone images processed with computer-vision algorithms. These datasets are analyzed using mathematical and statistical methods to produce digital terrain models and forest characteristics. (Iglhaut et al., 2019; Li et al., 2021)

In this work, aerial images from two test sites were collected with two different cameras and subsequently utilized in generating point clouds. Trees were evaluated by visual inspection of the point cloud and by using an open-source tree detection algorithm. The results were compared with measured sample trees.

The AI training of automated forest machine (AIFor) -project, which is related to the Smart, digitalized components and systems for data-based Agriculture and Forestry (AGRARSENSE) project,, aims to develop these new methods and a virtual environment for autonomous forest machine AI training. The AIFor-project is implemented by Lapland University of Applied Sciences, and it’s co-financed by Business Finland. The AGRARSENSE project was funded by the EU Chips Joint Undertaking.

Data collection

Two forest areas, “Hirvas 1” and “Hirvas 2” near Rovaniemi, were chosen for the study. “Hirvas 2” was imaged with two cameras to examine whether image quality affects the outcome of the forest simulation. The software used for point cloud generation was Agisoft Metashape (Agisoft Metashape, 2026), and the visual evaluation of the original and processed point cloud data was performed using the open-source software CloudCompare (CloudCompare, 2026).

Case A: Hirvas 1, Geodrone 4XL with SonyA6000 camera

The area was imaged in September 2023. The details of the camera and point cloud generation are shown in Table 1. The top view of the point cloud is shown in Figure 1 (left). For comparison, an aerial view of the same area is shown (image source: National Land Survey, processed with QGIS).

Table 1: Imaging and data processing details for case A.

Figure 1. Left: point cloud of case A; right: aerial image of Hirvas 1.

Case B: “Hirvas 2” Geodrone 4XL with SonyA6000 camera

The area was imaged in October 2024. The camera was the same as in case A, and other details are in Table 2. The general view of the area as seen in the point cloud is shown in Figure 2. 

Table 2: Imaging and data processing details for case B.

Case C: “Hirvas 2”, DJI Mavic 3 Enterprise with built-in camera 

The area was imaged in October 2024. The details of the camera and point cloud generation are shown in Table 3. The height point density in the point cloud results in a more detailed view of the environment. In Figure 2, a small section of the forest is shown in comparison with case B. 

Table 3: Imaging and data processing details for case C.

Figure 2. Left: general view of the point cloud generated from “Hirvas 2” (case B). Right: comparison of the point clouds generated with two different setups from the same forest area—case B on the left and case C on the right.

Comparison of sample trees with trees observed in point clouds and detected automatically

The point clouds were used for evaluating the approximate location and the height distribution of the trees, and the proportions of tree species. First, the point cloud was compared with measured data from the forest. The position, height and species of sample trees were collected. The position was measured using a hand-held GPS device. Also, sample plots from both sites were analyzed to obtain the species and tree height distribution on the forest area.

The measured trees were located in the point cloud based on their GPS coordinates. The tree height was roughly evaluated from the point cloud coordinates at the canopy and on the ground level near the tree. For each tree, it was also evaluated whether it was possible to see the tree species from the point cloud.

Moreover, a list of locations and heights of individual trees for each test case was obtained with a tree detection algorithm IndividualTreeDetection by the open-source toolkit WhiteBoxTools (Whitebox Geospatial Inc., 2026). The point clouds were first tiled to smaller section with CloudCompare, and the tiles were smoothed with top hat transformation as is suggested for the tree detection algorithm. The properties of the detected trees were also compared with the measured data.

Case A: “Hirvas 1” 

Nine trees from “Hirvas 1” were measured in May 2024. The observed properties of these trees were compared with those derived from the point cloud. The sample site measurements showed the tree species distribution of approximately 60 % pines, 30 % spruces, and 10 % birches. The average height of the trees was 14.50 m. The tree density was 1000 trees/ha. 

The heights of the trees evaluated from the point cloud were lower than those measured on-site. The tree heights were evaluated from both the original and top hat transformed point cloud by comparing the local maxima to the nearby ground points. The heights obtained from the individual tree detection algorithm were approximately the same as those evaluated directly from the point cloud. A possible source for the height error is that the ground under the dense canopies may not have been visible to the camera, and thus the lowest points in the point cloud are parts of tree branches.

It was not possible to visually distinguish the species of individual trees from the point cloud data top view, except for spruces due to their distinctive shape. The species of none of the trees was recognizable from top view of the point cloud.

The average height of at least 6 m tall trees found with the detection algorithm was 12.10 meters. The density of the trees found by the algorithm was obtained by computing all the detected trees within a radius of 20 m from a chosen point. The density of at least 6 m tall trees was 539 trees/ha, which is only 54 % of the density measured on-site. 

Moreover, it was noted that the individual tree detection algorithm identifies the tops of spruces correctly as distinct trees, but more branched species such as birches may exhibit multiple top points per tree. It is challenging to visually discern from the point cloud data whether a branched canopy belongs to a single tree or multiple trees, as the trunks are not visible.  

Case B: “Hirvas 2” (setup 1), and case C: “Hirvas 2” (setup 2) 

Measurements on two sample sites at “Hirvas 2” were done in November 2024. The features of the sites are given in Table 4.  Additionally, the coordinates, diameter at breast height, total height, and branch line were measured from 11 trees. 

Table 4. Sample sites analyzed at “Hirvas 2”. 

Site Density
(trees/ha) 		Tree proportions (%) Average tree height (m) 
1 1400 pine 50, spruce 7, birch 43 total 13.5, pine 16.3, spruce 12.0, birch 10.5
2 1300 pine 31, spruce, birch 31 total 10.8, pine 14.1, spruce 8.8, birch 10.1

Again, the visual evaluation of the tree species from the point cloud was not possible for most of the trees. In the point cloud of case C, the differences in the canopy shade and shape of the trees were more visible, and most of the tree species could be evaluated visually. Even the top view would be useful in classification by species. An example of trees as viewed from the side and top is shown in Figure 3.

Figure 3. An example of a side view (left) and top view (right) of the point cloud (case C).

A linear regression (Figure 4) was tested for the measured and evaluated tree heights. In case B, the correlation was not high (R2=0.49). In case C, a linear fit with R2=0.95 was found. The linear model could possibly be used in generating digital models; however, it should be calibrated for each setup.

Figure 4. Linear fit of tree height evaluated from point clouds and measured on-site.

The tree density observed at the sample site in case B was compared with the density of the trees found by the tree detection algorithm on either a larger rectangular forest area, or a polygonal forest area generated by k-means clustering (Figure 5). The detected density of 216-264 trees/hectares is significantly lower than 1300-1400 trees/hectare obtained from the sample sites. However, the area with lower than average tree density or missing on the border of the polygon may affect the computed density.

Figure 5. A polygonal forest area generated by k-means clustering (case B).

In case C, only a small sample was chosen from the point cloud for tree detection due to the large size of the data. The locations and heights of the trees are shown in Figure 6. The average height of the trees was 16.4 m. The density of detected trees on a rectangular area of approximately 118 m × 76 m was only 253 trees/hectare. Thus, even though the visual quality of the trees was better in this point cloud, the number of the detected trees does not correspond to the measured density.

Figure 6. Left: Selected area for tree detection from point cloud of case C, bordered with the yellow line. Background image is the point cloud of case B. Right: Tree locations and heights given by the tree detection algorithm, case C. 

The tree height distributions in cases B and C differ from each other as shown in Figure 7. In forest science, the height distribution of trees is often modeled with a lognormal function using the mean and standard deviation of tree height. Figure 7 (dashed lines) shows that the height distribution in both cases B and C follows a lognormal function. In both cases, the standard deviation is the same as observed in the sample site measurements. In case B, the meaning in the lognormal function is the same as the mean height of all trees on sample sites 1 and 2. In case C, a larger mean value (16.4 m) results in a better correspondence between the lognormal and detected distributions.

Figure 7. Cumulative proportions of trees belonging to different height classes in cases B and C. 

Methods for identifying the tree species

Even though identifying individual trees from visual data was challenging, we tested approaches for identifying the species of trees detected. First, the canopy shade was analyzed by computing the average red (R), green (G), and blue (B) components of points below and around detected treetops in case A. The average treetop shade was then compared visually with the point cloud data (Figure 8). Yellow shades were clearly visible in some areas of the point of cloud, likely representing deciduous trees, as the aerial images were recorded during autumn. This pattern was also observed in the clustering results.

Figure 8. Left: average color near the detected treetops; right: original data and treetops.

The RGB values were further analyzed by calculating the ratio of green and red components from the average G and R values around the detected treetops.  This G/R ratio could be used for distinguishing “pure green” and “greenish yellow” canopies. In Figure 9, red points indicate a high G/R ratio, and blue points indicate a lower G/R ratio. The difference is clearly visible in the point cloud. This information could be useful in evaluating the tree species distribution in the forest. 

Figure 9. Left: treetops labelled as “pure green” (red labels) and “greenish yellow” (blue labels). Right: Point cloud from the same area, showing the actual shades.

Moreover, analysis of the canopy shapes from the point cloud was attempted by computing the numbers of points in the point cloud near the detected treetop, thus trying to distinguish trees with a relatively narrow canopy, and trees with a wider canopy. Examining the original data visually (Figure 10) indicates the combined with the color analysis, it might be possible to evaluate an approximation for the tree species distribution. 

Figure 10. An example of a shape-based approach to tree species identification. The colors represent a numerical value characterizing the shape of the treetop: red shades corresponding to a narrow canopy, and yellow shades corresponding to a broad canopy. 

Summary and conclusions

When generating a digital twin on the forest, the proportions of different tree species, their height distributions, and tree density in the forest area must be known. The point cloud data showed – either directly or when used in a tree detection algorithm – in almost all cases too short trees, so calibration between the data and tree heights is necessary.

Identifying tree species from the point cloud visually can be challenging if the data is not very dense. However, methods of machine learning can be used for classifying the shades of the canopies. In this study, spruces, pines and birches (in autumn foliage) had distinctive shades that were detectable both visually and by clustering.

The tree density obtained with tree-detection algorithm was only about 25% – 50% of the measured density. On the other hand, trees with broad canopies could be detected as multiple trees. The tree detection algorithm works with specified minimum and maximum search radii, and probably different values work best for trees with broad or narrow canopies. Future work is needed to find optimal values for both cases. Moreover, the shape of the canopy can be utilized in evaluating the tree species when combined with the shade of the canopy.

The economic value of the forest, as well as the total biomass needed in ecological research, requires the breast-height diameters of the trees. As the point cloud is formed using aerial images, the trunks of the trees are not visible, thus the diameters cannot be evaluated from point clouds.

Sources

Agisoft Metashape, 2026. Referenced 19.5.2026. https://www.agisoftmetashape.com/

CloudCompare, 2026. Referenced 19.5.2026. https://www.cloudcompare.org/

Iglhaut, J., Cabo, C., Puliti, S., Piermattei, L., O’Connor, J., & Rosette, J. (2019). Structure from motion photogrammetry in forestry: A review. Current Forestry Reports, 5, 155–168. https://doi.org/10.1007/s40725-019-00094-3

Li, L., Wang, R., & Zhang, X. (2021). A tutorial review on point cloud registrations: Principle, classification, comparison, and technology challenges. Mathematical Problems in Engineering, 2021, 9953910. https://doi.org/10.1155/2021/9953910

Maanmittauslaitos, 2026. Referenced 19.5.2026. https://www.maanmittauslaitos.fi/en/e-services/mapsite

Murtiyoso, A., Holm, S., Riihimäki, H., Krucher, A., Griess, H., Griess, V. C., & Schweier, J. (2023). Virtual forests: A review on emerging questions in the use and application of 3D data in forestry. International Journal of Forest Engineering, 35(1), 29–42. https://doi.org/10.1080/14942119.2023.2217065

Visser, R., & Obi, O. F. (2020). Automation and robotics in forest harvesting operations. Croatian Journal of Forest Engineering, 42(1), 13–24. https://doi.org/10.5552/crojfe.2021.739

Whitebox Geospatial Inc., 2026. Referenced 19.5.2026. https://www.whiteboxgeo.com

From digital twin to AI-based perception system validation

Ramana Reddy Avula, Aleksi Narkilahti & Jussi Unga

Introduction

As part of use case 5 (Forestry) within the Smart, digitalized components and systems for data-based Agriculture and Forestry (AGRARSENSE) and AI training of automated forest machine (AIFor) projects, WP4 focused on transforming the simulator platform into a practical tool for safety-critical system validation. While WP2 concentrated on building the simulator architecture and sensor models, WP4 extended this work by creating a detailed digital twin of a real forest site in Vindeln, Sweden, and using it as the basis for high-fidelity data generation, virtual testing, and AI-driven perception research relevant to autonomous forestry machinery.

The Vindeln digital twin was developed using a combination of drone imagery, laser-scanning data, and on-site reference material collected together with project partners. This enabled accurate reconstruction of terrain, vegetation, and forest structures, allowing the simulator to reproduce realistic lighting conditions, seasonal variation, and occlusion patterns typical of boreal forests. The digital twin provided a controlled yet realistic environment in which simulated sensor outputs could be compared against real-world measurements and used for repeatable experimentation.

Within this environment, WP4 addressed three closely connected activities. First, the validated RGB camera model was evaluated in realistic outdoor conditions as part of the integrated system, complementing the controlled laboratory validation performed in WP2. Second, the digital twin was used to generate SimForest, an RGB-D instance segmentation dataset supporting the training and benchmarking of perception models for forestry applications (SimForest, 2026). Third, the simulator was employed as a virtual testing platform for safety-critical AI functionality, with a particular focus on human detection and emergency braking in autonomous forestry operations.

Together, these activities demonstrate how a simulation platform can move beyond pure development tooling and serve as a practical instrument for dataset generation, system-level validation, and scenario-based safety evaluation, supporting the development of future autonomous forestry machines. A short video demonstration of the simulator platform and forestry use case is available online (AGRARSENSE | Simulation in the Forestry Sector video, 2026).

Vindeln digital twin creation

The development of the Vindeln digital twin began with on-site data collection at the selected demonstration area in Vindeln, northern Sweden. The site was chosen together with project partners as a representative forestry environment suitable for evaluating autonomous transport and safety-critical perception functions. During the site visit in September 2024, extensive reference material was collected, including photographs, videos, drone imagery, and high-resolution laser scans of selected areas of interest. Particular attention was given to capturing both typical forest structures and distinctive landmarks, such as large boulders and dense vegetation, which are relevant for perception and obstacle-detection tasks.

During the site visit, the area was first walked through together with the lead partner to gain an overview of the environment and identify a suitable location for detailed scanning. A representative forest area containing a large boulder was selected as the primary focus for perception and validation experiments (Figure 1). Detailed data collection was then carried out, including photography, video recording, and laser scanning of the selected area using a Leica BLK360 scanner. In total, 30 scans were captured and later processed using Leica’s software to generate a high-resolution point cloud (Figure 2), which was imported into the simulator and used as a spatial reference for constructing the Vindeln digital twin and accurately placing terrain features, vegetation, and objects.

Figure 1. Representative forest location in Vindeln containing a large boulder, used as a reference object for perception and validation tasks.

Figure 2. Overview of the Leica BLK360 laser‑scanned point cloud of the selected area at Vindeln.

Following the on-site data collection, the broader Vindeln environment was constructed using drone-derived heightmap and canopy height model (CHM) data provided by project partners. The landscape was generated from the heightmap and cross-referenced with locally captured point-cloud data and open-source road information to ensure correct scale and alignment (Figure 3). Compared to earlier work in the Hirvas area, the Vindeln heightmap data enabled a higher-resolution landscape with more accurate height variations and improved tree placement, providing a stronger basis for realistic simulation.

After initial terrain generation, remaining height inconsistencies were smoothed, and the detailed Vindeln area was blended into the surrounding terrain to achieve a natural transition and a coherent sense of scale. Road placement and geometry were verified using OpenStreetMap data imported via the Blender GIS plugin and aligned within Unreal Engine. To place vegetation, the CHM data was processed in QGIS and converted into a color-encoded height map, which was then used to position appropriately sized trees with an approximate species distribution (Figure 3). Smaller environmental elements such as rocks, stumps, ground vegetation, and forest debris were procedurally scattered to enhance realism, with landscape layers used to prevent vegetation placement in unsuitable areas such as bogs and wetlands.

Figure 3. Generated Vindeln digital twin area. Left: Terrain generated using high‑resolution heightmap data with accurately aligned road geometry. Right: Vegetation placement based on canopy height model (CHM) data, enabling realistic tree heights and distribution.

In addition, a smaller subsection of the Vindeln area was modelled directly from the Leica BLK360 point-cloud data, providing highly accurate tree heights and object positions, and was manually reconstructed in the simulator to closely match the real-world scene (Figure 4). This laser-scanned region, together with the images captured on site, served as a high-fidelity reference for validating and aligning the procedurally generated parts of the digital twin, ensuring that the final environment closely reflects the real-world Vindeln forest site.

Figure 4. Hand-placed area in Vindeln using point-cloud data

Camera-based perception validation in the Vindeln digital twin

With the Vindeln digital twin established, WP4 extended the camera model validation beyond controlled laboratory conditions into realistic forest scenarios. While WP2 focused on validating the RGB camera model in an isolated, reproducible laboratory setup, WP4 evaluated how the same camera model behaves when immersed in a complex outdoor environment reconstructed from real-world sensor data. This allowed the validation process to capture the interaction between lighting, vegetation, terrain, and simulated sensor behavior, which is difficult to fully reproduce indoors.

To enable direct comparison with the simulated results, real-world reference data were collected at the actual forest site in Vindeln. On site, Color Checker images and spectrometer readings were recorded under naturally cloudy autumn conditions. These measurements were then aligned with the corresponding scenes in the Vindeln digital twin, allowing a thorough comparison of color fidelity across seasons, weather variations, and vegetation states. Figure 5 illustrates how seasonal rendering and environmental conditions influence the visual output of the simulated camera.

Figure 5. Comparison of camera outputs across seasons during scenario-based tests. From left to right: real‑world autumn reference, simulated autumn, simulated summer, and simulated winter.

The analysis showed that while most simulated seasonal conditions produced ΔE color difference values within an acceptable range, the simulated autumn environment exhibited noticeably higher deviations compared to real-world images. This discrepancy was linked not to the camera model itself but to environmental rendering limitations; specifically, the lack of realistic leaf‑shedding behavior and non-uniform autumn foliage colors. As a result, a new functional requirement was introduced to support configurable leaf‑color variation, and the vegetation models were refined accordingly. These updates improved the digital twin’s seasonal realism and enhanced the accuracy of subsequent perception tests.

SimForest dataset

Building on the Vindeln digital twin, a high-fidelity synthetic dataset was also generated to support perception research for autonomous forestry machinery. This effort resulted in SimForest, a 4K-resolution RGB-D instance segmentation dataset created entirely within the Unreal Engine 5-based forestry simulator (Unreal Engine 5, 2026). The dataset contains 5,000 images, each consisting of photorealistic RGB frames, 32-bit depth maps encoded as RGB PNG images, and instance segmentation masks for all objects within 15 meters of the camera (SimForest 2026). Figure 6 presents representative samples from the dataset. Each image is accompanied by extensive metadata, including camera intrinsics, camera poses, environmental parameters such as season, time of day, month, and cloudiness, as well as object-level information such as 3D position, orientation, and physical size.

Figure 6. Samples from the SimForest dataset showing aligned RGB images, scene depth maps, instance segmentation masks, and terrain depth maps under diverse environmental conditions. Each column represents a different data sample.

SimForest was generated by capturing images from synchronized virtual RGB, depth, and instance‑segmentation cameras configured at 3840×2160 resolution with a 90° horizontal field of view. A 200 m × 200 m area of the Vindeln digital twin was sampled using a regular grid, with each grid cell containing a camera viewpoint placed 2 meters above ground and oriented using one of 16 randomized yaw angles. Environmental conditions were randomized using real solar angle calculations based on geolocation, producing diverse lighting and seasonal variations that help improve the robustness of trained models.

Across all images, SimForest contains 40,554 instance annotations covering 11 object categories, heavily dominated by pine and spruce trunks and foliage. Each annotated object includes a pixel‑accurate instance mask, a 2D bounding box, an estimated 3D bounding box derived from the depth map, and complete instance metadata. A quality‑control step removes frames with extreme occlusion or limited visibility to ensure the dataset remains suitable for machine‑learning workflows. Models trained solely on SimForest demonstrated strong performance: YOLOv11‑based experiments achieved mAP@50 of 0.92 for trunk detection and 0.74 for instance segmentation, confirming the dataset’s ability to support near‑field RGB‑D perception in complex forest environments. Because the dataset is explicitly tied to the Vindeln digital twin and its real‑world counterpart, it also provides a strong basis for future research on sim‑to‑real transfer and validation of AI‑based safety‑critical functions in autonomous forestry machines.

Using the simulator for virtual testing and validation

With the digital twin and sensor models in place, WP4 also used the simulator as a platform for virtual testing and validation of safety‑critical functions in autonomous forestry machinery. The focus of this work was the human detection and emergency braking functions, where the simulator served as a controlled environment for systematically assessing how an autonomous forestry shuttle responds when a person enters its forward hazard zone. As illustrated in Figure 7, the hazard zone is defined as an arc extending ahead of the vehicle, sized according to the maximum stopping distance and the sensor field of view, and is associated with a primary safety goal (SG1) and three Safety Performance Indicators (SPIs). Entry of a person into this zone represents a critical safety risk, requiring the system to reliably detect humans and trigger an immediate transition to a safe stop.

Figure 7. Operational context of the AGRARSENSE forestry shuttle, illustrating the primary safety goal (SG1) and the three associated Safety Performance Indicators (SPIs).

To support this assessment, the simulator reproduced the shuttle’s intended operating conditions: low‑speed travel along mapped forest roads, limited visibility, variable terrain, and the presence of dynamic obstacles such as human operators. Using the virtual RGB and depth cameras mounted on the shuttle and a lightweight YOLO‑based detector, the simulator enabled repeatable and scalable validation runs. Each scenario began with the shuttle moving along a geofenced corridor, after which a human model entered the forward hazard zone at a predefined moment. The shuttle’s perception system, integrated through the WayWiseR platform, monitored the hazard zone continuously and issued an emergency‑stop command when detection occurred or when perception confidence fell below acceptable thresholds.

A series of 50 simulation runs were executed under clear daytime conditions, each with the shuttle travelling at approximately 5 km/h and a simulated human model placed along its path. Across all runs, the perception system achieved a 0% false‑negative rate, demonstrating consistent detection even under variations in lighting, viewpoint, and scene geometry. Reaction times from detection to the emergency‑stop command had a mean of 73.16 ms, and after braking, the mean distance margin between the shuttle and the human model was 12.11 m, indicating compliance with the defined safety goal of preventing collision within the hazard zone. The simulator thus provided quantitative evidence of the system’s capability to detect a person and initiate braking with sufficient margin, while allowing safe and repeatable evaluation of edge cases that would be challenging or unsafe to reproduce physically.

These results demonstrate the value of the AGRARSENSE simulator as a virtual testing tool for safety‑critical AI functionality. By enabling controlled scenario execution, the simulator supports systematic validation of perception and emergency‑response behavior before physical testing. This strengthens confidence in the system’s robustness and provides a scalable pathway for future validation activities involving more diverse environmental conditions and degraded sensing scenarios.

Summary

Overall, the WP4 work was successfully completed, resulting in a functional and versatile simulator platform capable of supporting realistic forestry scenarios. A high-fidelity digital twin of the Vindeln forest site was created based on detailed field measurements and remote-sensing data, enabling accurate representation of terrain, vegetation, and seasonal environmental conditions. This digital twin formed the foundation for all subsequent WP4 activities.

Within this environment, the simulator was used to integrate and evaluate the RGB camera model under realistic outdoor conditions, complementing the laboratory-based sensor validation carried out in WP2. WP4 also produced the SimForest RGB-D instance segmentation dataset, consisting of 5,000 high-resolution annotated images with aligned depth, instance labels, and rich metadata, providing a valuable resource for training and benchmarking perception models for forestry applications (SimForest, 2026).

In addition, the simulator was employed as a virtual testing platform for safety-critical AI functionality, with scenario-based evaluation of human detection and emergency braking behavior in an autonomous forestry shuttle. These tests demonstrated reliable detection performance, fast reaction times, and safe stopping behavior under controlled yet realistic conditions, highlighting the simulator’s value for systematic validation of safety-critical functions.

Throughout WP4, close collaboration with UC5 partners ensured that the simulator addressed practical requirements and supported partner use cases. Feedback from these partners guided continuous refinement of the platform. Overall, WP4 demonstrates how the AGRARSENSE simulator can be used not only as a development tool, but also as a credible environment for dataset generation, integrated sensor validation, and virtual testing of autonomous forestry systems. The simulator is expected to continue supporting research and development activities beyond the project.

The AIFor project, which is related to the AGRARSENSE project, aims to develop new methods and a virtual environment for autonomous forest machine AI training. The AIFor project is implemented by Lapland University of Applied Sciences and co-financed by Business Finland. The AGRARSENSE project was funded by the EU Chips Joint Undertaking.

Sources

AGRARSENSE 2026. Referenced 16.4.2026. https://www.agrarsense.eu/

SimForest: RGB-D Instance Segmentation Dataset, 2025. Referenced 15.5.2026. https://ieeexplore.ieee.org/document/11396858

AGRARSENSE | Simulation in the Forestry Sector video, 2025. Referenced 16.4.2026.

Oho! Tämä YouTube-upotus ei näy, koska et ole hyväksynyt markkinointievästeitä. Hyväksy markkinointievästeet.

Unreal Engine 5, 2026. Referenced 15.5.2026. https://www.unrealengine.com/

Building a Foundation for Future Autonomous Systems

Anne Saloniemi

Lapland University of Applied Sciences (Lapland UAS) participated in the Smart, digitalized components and systems for data-based Agriculture and Forestry (AGRARSENSE) and AI training for automated forest machine (AIFor) iniatives during January 2023 and February 2026. These projects have generated a set of complementary results that contribute to the advancement of digital, AI-based solutions for forestry and autonomous off-road machinery.

The key outcomes include a high-fidelity forestry simulator platform (Figure 1), validated virtual sensor models, digital twins of real forest environments, synthetic datasets for machine learning research, and practical experience in applying simulation-based methods to safety critical system validation. These assets provide value far beyond the initial development phase, supporting research, education, and industrial innovation.

Figure 1. Forestry Simulator Platform. FrostBit Software Lab.

Exploitation within research and development

One of the key exploitable outcomes is the Unreal Engine–based forestry simulator developed at Lapland UAS in collaboration with project partners. The simulator provides a flexible platform for continued research on autonomous driving, perception, and decision making in complex forest environments. The developed digital twins of the Hirvas and Vindeln forest sites, together with the associated workflows for terrain, vegetation, and environmental modelling, form reusable assets for ongoing RDI activities.

As an openly published dataset, it can be reused by researchers and developers for training, benchmarking, and comparative studies of AI‑based perception methods. The dataset also supports future work. Industrial partners and technology providers can exploit the project results by using the simulator and digital twins as test environments for software components, sensors, and AI models.

Academic Contributions and Resources:

  • SimForest: RGBD Instance Segmentation Dataset   (https://ieeexplore.ieee.org/document/11396858)
  • BRINGING FOREST GROWTH MODELS INTO THE CLASSROOM: ACCESSIBLE SIMULATION TOOLS FOR TEACHING AND RESEARCH (https://zenodo.org/records/18469317)
  • Generative AI and Simulation-Based Data Augmentation for Enhanced Object Detection in Low-Data Forestry Environments (https://www.mdpi.com/1999-4907/17/3/302)

Demonstrator Video:

  • AGRARSENSE  2025 demonstrator video of Forestry Use Case (https://zenodo.org/records/17975164)

Insights from the ”Pohjoisen tekijät” Blog

  • Making realistic 3D trees. https://lapinamk.fi/en/blogArticle/making-realistic-3d-trees/  
  • Developing a simulator platform to advance autonomous systems in forestry usage. https://lapinamk.fi/blogiartikkeli/developing-a-simulator-platform-to-advance-autonomous-systems-in-forestry-usage/
  • Perinteiset metsänmittausmenetelmät tukevat metsätalouden digitalisaatiota https://lapinamk.fi/blogiartikkeli/perinteiset-metsanmittausmenetelmat-tukevat-metsatalouden-digitalisaatiota/#:~:text=Mets%C3%A4talousinsin%C3%B6%C3%B6ri%20(YAMK)%20Veli%2DPekka,laserkeilausten%20avulla%20sek%C3%A4%20t%C3%A4ydent%C3%A4en%20satelliittikuvilla.

Thesis Work:

  • Synteettisen aineiston käyttö kuvantunnistuksessa https://www.theseus.fi/handle/10024/913958

Impact on Education and Competence Development

The simulator, digital twins, and datasets are well suited for use in higher education and professional training at Lapland UAS. By using real project outputs and case studies, teaching can better bridge theory and practice, strengthen students’ applied skills, and improve their readiness for working life.

Future use and sustainability of results

The exploitation of AGRARSENSE and AIFor results does not end with this publication. Lapland UAS and its partners aim to continue using and further developing the simulator platform, digital twins, and datasets in future projects, education, and collaboration with industry.

The continuity of this work is ensured in several ways:

  • New Projects: The methods learned will be used in future national and international research projects.
  • Collaboration with Industry: The simulator provides companies with a safe environment to test new software and equipment before installing them on real machines. This speeds up the creation of new solutions.
  • Open Information: By sharing findings and data openly, it is possible for others to learn and develop new ideas based on what has been achieved in these projects.
  • Benefits to Other Fields: Although this work focused on forestry, the same methods can be used in other sectors, such as agriculture, mining, or logistics.

Overall, the projects demonstrate how simulation, digital twins, and AI can be combined to support safer, more efficient, and more sustainable forestry operations. The results provide a strong foundation for continued research, innovation, and skills development in autonomous offroad systems and offer transferable solutions applicable far beyond the original project context.