Research
Our research interests span from Embodied AI and Robotics to Physical Simulations and Digital Twins:
Digital Twins for Wildfires
A major direction of our research is the modeling and understanding of fire. This research is supported by through an ERC Consolidator grant. More details of this project are available here: www.wildfiretwins.org.
Wildfires are one of the most destructive natural disasters. Their unpredictable behaviour and complex physical dynamics make them incredibly challenging to manage. Traditional methods often fall short in keeping up with their fast-moving nature, creating a need for more effective solutions. Digital twins, which combine virtual models and real-world simulations, have emerged as a promising tool to address this issue. With this in mind, the ERC-funded WildfireTwins project aims to create detailed 3D models of ecosystems, coupled with physical simulations, to develop real-time wildfire simulations. These digital twins will enable better decision-making for firefighting services. By using photorealistic imagery and AI-driven tools, the project will offer innovative solutions for combating this global threat.
Selected Publications:
-
J. Nazarenus, Dominik Michels, W. Palubicki, S. Kou, F.-L. Zhang, S. Pirk, R. Koch, Gaussians on Fire: High-Frequency Reconstruction of Flames, 2025, [Arxiv], [Preprint]
-
H. Wrede, A. R. Wagner, S. M. Mahfuz, W. Pałubicki, D. L. Michels, S. Pirk, Fire-X: Extinguishing Fire with Stoichiometric Heat Release, ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia), 2025, [Website], [Preprint], [Video]
-
D. Liu, J. Klein, W. Pałubicki, S. Pirk, D. L. Michels, Flame Forge: Combustion of Generalized Wooden Structures, ACM SIGGRAPH / Eurographics Symposium on Computer Animation (SCA), 2025, [Preprint], [DOI], [ArXiv], [Video]
-
A. Kokosza, H. Wrede, D. G. Esparza, M. Makowski, D. Liu, D. L. Michels, S. Pirk, W. Pałubicki, Scintilla: Simulating Combustible Vegetation for Wildfires, ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2024, [Website], [Preprint], [Video], [Bibtex]
-
T. Hädrich, D. T. Banuti, W. Pałubicki, S. Pirk, D. L. Michels, Fire in Paradise: Mesoscale Simulation of Wildfires, ACM Transactions on Graphics (SIGGRAPH), 2021, [Website], [Preprint], [Video], [Talk], [Two Minute Papers], [Gallery of Fluid Motion], [Bibtex]
Embodied AI and Robotics
A major direction of our research is to develop AI models for Embodied AI and Robotics. This research is supported by our Interreg research project wit partners from the Southern Demark University, University of Lübeck, the University of Applied Sciences Flensburg, DFKI, Harting and Novo Nordisk: link.
Agent-centric representations in embodied artificial intelligence integrate perception, action, and learning by grounding intelligence in an agent’s physical body and its continuous interaction with the world. In the context of AI, machine learning, and robotics, these representations prioritize the agent’s own sensory inputs, motor capabilities, internal state, and goals as they unfold within a dynamic environment. This perspective stands in contrast to environment-centric or third-person models, shifting the focus toward first-person, embodied experience and closed-loop perception–action cycles.
The objective of agent-centric representations in embodied AI is to enable agents to understand, reason about, and adapt to their surroundings through direct interaction -- learning from their actions, consequences, and constraints imposed by their embodiment. By explicitly modeling how an agent’s body, sensors, and actuators shape its experience of the world, these representations support more robust decision-making, generalization, and autonomy in real-world settings.
Selected Publications:
-
A. Francis, C. Pérez-D' Arpino, C. Li, F. Xia, A. Alahi, R. Alami, A. Bera, A. Biswas, J. Biswas, R. Chandra, H.-T. L. Chiang, M. Everett, S. Ha, J. Hart, J. P. How, H. Karnan, T.-W. E. Lee, L. J. Manso, R. Mirksy, S. Pirk, P. T. Singamaneni, P. Stone, A. V. Taylor, P. Trautman, N. Tsoi, M. Vázquez, X. Xiao, P. Zu, N. Yokoyama, A. Toshev, R. Martín-Martín, Principles and Guidelines for Evaluating Social Robot Navigation Algorithms, ACM Transactions On Human-Robot Interaction, 2025, [Preprint], [ArXiv], [DOI]
-
C. Cuan, E. Lee, E. Fisher, A. Francis, L. Takayama, T. Zhang, A. Toshev, S. Pirk, Gesture2Path: Imitation Learning for Gesture-aware Navigation, International Conference on Social Robotics (ICSR), 2024, [Preprint], [ArXiv], [Bibtex]
-
A. Francis, C. Pérez-D' Arpino, C. Li, F. Xia, A. Alahi, A. Bera, A. Biswas, J. Biswas, H.-T. Lewis Chiang, M. Everett, S. Ha, J. Hart, H. Karnan, T.-W. E Lee, L. J. Manso, R. Mirksy, S. Pirk, P. T. Singamaneni, P. Stone, A. V. Taylor, P. Trautman, N. Tsoi, M. Vázquez, X. Xiao, P. Xu, N. Yokoyama, R. Martín-Martín, A. Toshev, Benchmarking Social Robot Navigation Across Academia and Industry, Best Paper Award Nominee, AAAI Spring Symposium, 2023, [Preprint], [Bibtex], [Symposium Website]
-
H. Karnan, A. Nair, X. Xiao, G. Warnell, S. Pirk, A. Toshev, J. Hart, J. Biswas, P. Stone, Socially Compliant Navigation Dataset (SCAND): A Large-Scale Dataset of Demonstrations for Social Navigation, IEEE Robotics and Automation Letters (RA-L) and IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, [Website], [Preprint], [Video], [Dataset], [Bibtex]
Generative AI
Generative AI focuses on models that can synthesize, reconstruct, and reason about complex data by learning underlying structure and variability. In our research, we explore generative methods and large reconstruction models (LRMs) that operate across images and geometry. These models support tasks such as image and shape synthesis, reconstruction, and neural mesh generation.
In medical AI, generative models are applied to produce accurate and clinically meaningful reconstructions from medical data. This includes medical image super-resolution for modalities such as CT, where generative approaches enhance spatial detail, as well as neural mesh generation for reconstructing anatomical structures. These representations enable patient-specific modeling and implant research, supporting personalized analysis, design, and simulation.
This work is conducted as part of the Interdisciplinary Center for Scientific Computing and Imaging (IDIR) at Kiel University (link).
Selected Publications:
-
N. C. Koser, M. J. Finck, F. N. von Brackel, B. Ondruschka, S. Pirk, C.-C. Glüer, Real Super-Resolution for Proximal Femur: Enhanced Computation of Structural Bone Metrics from Clinical CTs, International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2025), [Preprint], [Poster]
-
N. C. Koser, M. J. Finck, J. B. Hövener, C. C. Glüer, S. Pirk, FemoraLyze: A Modular Framework for Proximal Femur Analysis, Medical Imaging with Deep Learning-Short Papers, 2025, [Website], [Preprint]
-
D. Xie, S. Bi, Z. Shu, K. Zhang, Z. Xu, Y. Zhou, S. Pirk, A. Kaufman, X. Sun, H. Tan, LRM-Zero: Training Large Reconstruction Models with Synthesized Data, Conference on Neural Information Processing Systems (NeurIPS), 2024, [Website], [ArXiv], [Preprint]
-
A. Maesumi, D. Hu, K. Saripalli, V. G. Kim, M. Fisher, S. Pirk, D. Ritchie, One Noise to Rule Them All: Learning a Unified Model of Spatially-Varying Noise Patterns, ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2024, [Website], [ArXiv], [Preprint], [Bibtex]
Environmental Modeling
Environmental modeling aims to represent the structure and dynamics of complex natural and man-madeenvironments, capturing how physical and biological processes evolve over time. This includes modeling weather systems, vegetation such as trees and forests, urban form, and interconnected ecosystems. These environments are inherently dynamic and multi-scale, shaped by interactions between atmosphere, terrain, living organisms, and human infrastructure, making accurate and efficient modeling a fundamental challenge.
The objective of environmental modeling is to enable predictive, interactive, and physics-based representations that support analysis, decision-making, and learning. By modeling phenomena such as weather dynamics, tree growt, and ecosystem processes, these approaches provide a foundation for understanding environmental change and for coupling environments with embodied agents. Such models are essential for studying climate impacts, urban sustainability, and ecological resilience, as well as for creating realistic, responsive worlds in which intelligent systems can perceive, act, and learn.
Selected Publications:
-
X. Zhou, B. Li, B. Benes, A. Habib, S. Fei, J. Shao, S. Pirk, TreeStructor: Forest Reconstruction with Neural Ranking, IEEE Transactions on Geoscience and Remote Sensing, 2025, [Website], [Preprint], [Video]
-
B. Li, N. Schwarz, W. Pałubicki, S. Pirk, B. Benes, Interactive Invigoration: Volumetric Modeling of Trees with Strands, ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2024, [Website], [Preprint], [Video], [Bibtex]
-
J. A. Amador Herrera, J. Klein, D. Liu, W. Pałubicki, S. Pirk, Dominik, L. Michels, Cyclogenesis: Simulating Hurricanes and Tornadoes,
ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2024, [Website], [Preprint], [Video], [Bibtex]
-
W. Pałubicki, M. Makowski, W. Gajda, T. Hädrich, D. L. Michels, S. Pirk, Ecoclimates: Climate-Response Modeling of Vegetation, ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2022, [Website], [Preprint], [Video], [Bibtex]