38. Lower Saxony Day

We would probably all like to have a robot at home that tidies up or takes out the garbage. But first it needs an idea of where it is, what is happening around it and what it means to be where it is. What we humans take for granted must first be taught to the robot. And in such a way that we continue to understand why it does what it does. This not only creates trust, but also makes it possible to correct its behavior in case of doubt.

The first step is to recognize its environment and what it can deduce from it in order to perform its task. An important data source for this are 3D laser scanners, which provide detailed information about the environment in the form of point clouds.

3D laser scans are of great importance for mobile robots because they enable precise recording of the environment in three dimensions. By subdividing this laser scan data, the robot can segment the point cloud into different areas - for example into floors, walls, obstacles or moving objects. This division helps the robot to better understand its environment, recognize relevant structures and move safely. Especially in complex or dynamic environments, the segmentation of 3D laser scans enables reliable navigation, obstacle avoidance and targeted interaction with the environment.

As mobile robots usually have limited computing resources, efficient processing of 3D data is crucial. The segmentation algorithms must not only be accurate, but also fast and resource-efficient in order to be executed in real time on the robot. This is the only way it can react quickly to changes in its environment and act reliably and autonomously.

The working group uses drone technology in numerous research projects to obtain a wide range of information about landscapes and plants. In recent years, there have been numerous collaborations in research projects in an agricultural context.

The use of multispectral drone technology was demonstrated in real time at the stand. The live image from a DJI Phantom4 multispectral was transmitted directly to a monitor. During the demonstration, the NDVI (Normalized Difference Vegetation Index) was calculated and visualized on-the-fly. This vegetation-specific index enables a precise assessment of plant condition and vitality.

Real and artificial vegetation were used to demonstrate how multispectral sensor technology provides information that remains hidden to the human eye. The technology opens up a wide range of potential applications, for example in agriculture or environmental monitoring.

In order to make AI methods usable on low-cost and therefore resource-limited systems, we are researching various approaches. First of all, it is important that the entire system development process is "resource-aware". To this end, we are working on prediction modelsthat can accurately estimatethe memory requirements, latency and throughput of an AI product line for combine harvesters for various hardware and software configurations, for example. In order to generate minimal AI models tailored to the use case, we are investigating methods to generate these automatically on the basis of large, complex models. For example, we were able to teach our robot dog LESSI voice commands without having to rely on data-intensive preliminary work such as the collection and annotation of thousands of training data. In addition, the application of trained models must be implemented extremely efficiently. To this end, we are developing so-called "executable" AI models, e.g. "executable neural networks", which are represented in memory not as a data structure but as a sequence of machine instructions. This minimizes the execution time.

Interactive elements are often used in didactics to explain technology clearly to students, pupils and other interested parties. The interactive Nim game, for example, demonstrates how machine learning works. The exhibit demonstrated how reinforcement learning works in principle.

In the lidar game, you explore an unknown environment with the help of a simulated lidar sensor. You see the world as a robot "sees" it, have to orient yourself in it and reach a goal.

The model drone simulates how a drone or satellite collects data and reconstructs a three-dimensional image of the earth's surface.

We are living in a time of great challenges (e.g. species extinction, environmental pollution, overuse of resources, land use, climate change) and impressive new opportunities (e.g. innovative technologies, artificial intelligence, efficiency improvements). This is particularly evident in the area of agriculture and food. Many of the examples mentioned are interconnected and influence each other. For sustainable development, it is therefore important not to look at complex problems in isolation, but to understand them in a system context - i.e. as individual parts interacting with the larger whole. Environmental systems science deals with these problems in an interdisciplinary manner and with special tools, namely systems analysis, mathematical models and computer-based simulations.