Our machine learning research focuses on data-driven methods for analyzing complex, often high-dimensional datasets - for example, terrestrial point clouds in geophysical exploration or time-series based air traffic forecasting. We combine supervised learning, deep learning, and hybrid approaches (e.g., LSTMs with evolutionary algorithms) to model and interpret patterns, anomalies, and temporal dynamics. A key focus is the development of model-driven ML pipelines that integrate heterogeneous data sources and enable automated reasoning - including the extraction of physically grounded prediction models from neural networks.

Our research also focuses on deploying ML models on resource-constrained platforms such as ESP32 microcontrollers, wearables, and mobile robots. By combining efficient model design with hardware-aware optimization and data preparation, we enable on-device inference where energy, memory, and processing power are limited. The result: autonomous systems that can make fast, reliable decisions directly at the edge, without constant reliance on remote servers.

Many application domains lack sufficient annotated depth sensor data for training modern AI systems. To address this gap, we develop methods for the automated generation of synthetic 3D data in virtual environments. Using game engines and custom tools like BlAInder, we create realistic, semantically annotated point clouds for sensors such as LiDAR and Sonar - including complex environmental effects like rain, fog, or underwater physics. These virtual sensors enable scalable, controlled, and repeatable data generation for AI development in areas like robotics, autonomous driving, and geophysics.

Using synthetic datasets, virtual LiDAR, and advanced simulation environments, we investigate how perception models can be trained and validated safely before being deployed in the real world. A key focus is bridging the sim-to-real gap, ensuring that algorithms developed in virtual spaces remain robust in dynamic, real environments. This research accelerates development cycles while reducing cost and risk.