Applied Artificial Intelligence

While our research is quite broad, as evidenced by the list above, a central theme in the group's research is interpretable AI, i.e., methods, processes, and systems that are inherently human-interpretable, thus offering transparency, accountability and safety. This is in stark contrast to the currently dominant trend in AI that instead is strongly focused on black-box (mainly neural) methods. On the positive side, black box methods have improved performance across many subfields of AI and also made entirely new applications (e.g., LLM-based chatbots) possible. On the other hand, black-box AI and, in particular, generative AI suffer from inherent limitations in terms of reliability and safety, as evidenced by many examples of catastrophic failures of black-box AI systems.

By contrast, in our research, we use so-called glass-box methods. In such approaches, it is possible for a human observer to understand the inner workings of the model. From a methodological standpoint, some of the core aims of our research are (i) to raise the performance of glass-box models so as to reach performance levels on a par with black-box methods, and (ii) to develop and apply general-purpose training approaches for glass-box methods.

Natural language processing

In this field, we have multiple ongoing projects. Our main focus areas are (i) text classification and (ii) interpretable language models. In the case of text classification, we have developed a
fully interpretable method that achieves state-of-the-art performance levels by using an approach inspired by the pre-training applied in black-box methods. In the case of language models, we are exploring various ways of building generative models that, unlike transformer-based LLMs, offer transparency and accountability.

Fleet coordination

In this project, we study automated coordination of vehicle fleets consisting of many vehicles. We have developed a general method for automated trajectory planning for this application.
The method handles both the path planning and, crucially, the timing of the vehicles' motions, thus guaranteeing efficient and collision-free operations. We have applied our method to situations involving automated mining vehicles, and are currently expanding our approach to the problem of automated railway shunting, in a case with battery-powered railcars capable of low-speed motion at shunting yards.

Autonomous Robot for Data-Driven Railway Maintenance

Within the IAM4RAIL project, in close collaboration with the Swedish Transport Administration (Trafikverket), we are developing autonomous robotic platforms for inspection and maintenance of railway infrastructure. The project's main objective is to carry out data-driven maintenance via robots that can autonomously navigate the rail environment, collect high-resolution condition data, perform inspections, and carry out minor maintenance interventions on site. The VIDAR robot is a central resource in the project that feeds real-time data to digital twins and predictive models, thereby contributing to the main goals of the project, namely, to increase safety, reduce traffic disruptions, and extend infrastructure lifespan.

Forestry and Geospatial AI

In forestry, our work focuses on the use of interpretable AI for analysing forest ecosystems from geospatial and remote sensing data. A central question is to what extent structural information observable from above, e.g., canopy height, spatial heterogeneity, and canopy gaps, can be used to infer ecological properties that are otherwise assessed through field surveys. In particular, we study how high-resolution canopy height models and related data sources can support the estimation of indicators such as forest naturalness and conservation value.

Monitoring and Governing Ocean health

Our work within the Mistra C2B2 project is dedicated to advancing data-driven innovation in Sweden's blue economy. We focus on integrating marine and atmospheric data from sectors such as fisheries, shipping, and offshore activities. By developing and using underwater data collection platforms, such as fixed observatories, surface drones, and gliders, the project aims to improve our understanding of marine ecosystems. Our key deliverables include an open inventory of ongoing marine projects and prototypes aimed at improving governance.