The dispersion of gases or substances in the air is a highly complex process due to atmospheric dynamics. Therefore, area-wide measurements of such processes are often very difficult or technically unfeasible. In many cases, emissions of carbon dioxide and methane, for example, are currently monitored by permanently installed measuring sensors or by portable measuring instruments. Thus, measurements are carried out at a few positions, which significantly limits the spatial resolution.
Although portable instruments can be used at landfills, they require additional people on site to obtain measurements. Hence, the development and research of autonomous mobile robotic systems and algorithms is considered key to the efficient and robust exploration of gaseous processes.
In contrast to imaging cameras, which offer high spatial and temporal resolution and record information from a certain distance, numerous non-visual sensors pose a particular challenge. On the one hand, such sensors often only allow point-by-point measurement; on the other hand, they also have low update rates and long impulse response times. This requires completely new approaches for autonomous robotic environment perception and gas sensing. The project STARE (Swarm-based Technologies for Autonomous Robotic Exploration in the Air and on the Ground) is therefore focusing on designing an AI-based, generic architecture concept for autonomous measurement and exploration tasks involving non-visual sensor data.
The project focuses on:
distributed data acquisition and processing using cooperative multi-agent systems (so-called swarms).
model-based machine learning, for which physical models of the spatial propagation of environmental parameters (essentially gases) are combined with machine learning (ML) approaches.
The generic architecture concept can be used for a variety of applications in disaster control, defence, environmental and planetary exploration. Compared to conventional solutions, the approach also makes it possible to significantly increase the spatial sampling resolution and reduce the amount of training data required.
