Teaching
Since 2023 I teach robotics and perception at IU International University of Applied Sciences in a fully online, self-paced distance-learning format. I supervise Bachelor's and Master's theses, with a growing focus on NVIDIA simulation and robot-learning tooling (Isaac Sim, Isaac Lab, Omniverse).
Courses
Six courses at Bachelor's and Master's level, delivered asynchronously through IU's distance-learning platform.
-
Mobile Robotics
Covers the full stack of autonomous mobile systems: locomotion, kinematics and dynamics, perception (LiDAR, IMU, vision, GPS), path and motion planning, and a short discourse on Simultaneous Localization and Mapping (SLAM).
-
Project: Applied Robotics with Robotic Platforms
Hands-on project course with a strong emphasis on ROS 2 as the de-facto middleware of modern robotics. Students move from theory to actually running code on simulated robots using an interactive online learning platform.
-
Computer Vision for Autonomous Systems
From the physics of image formation to modern deep learning perception pipelines: camera models and calibration, multi-sensor architectures, image processing, feature detection, object detection and tracking, semantic and instance segmentation.
-
Architectures of Self-Driving Vehicles
A full-stack view of the autonomous vehicle architecture: environment perception, sensor fusion and state estimation, vehicle dynamics with longitudinal and lateral control, vehicle-to-X communication, and the social and ethical implications of self-driving technology.
-
Case Study: Localization, Motion Planning and Sensor Fusion
Hands-on case study course applying localization, motion planning, and sensor fusion methods to two concrete autonomous vehicle scenarios: a self-driving car on the road and an autonomous system in a manufacturing facility.
-
Soft Robotics
An introduction to compliant, deformable robotic systems inspired by biological structures. Covers soft actuator principles, sensing for soft robots, modeling of continuum structures, and control strategies.
Supervised theses
12 theses as first supervisor and 10 as second supervisor at IU (2024–2026).
-
Benchmarking Physics-Informed Reinforcement Learning Across Simulated Physical Domains
Design of a standardized, reproducible evaluation framework for Physics-Informed Reinforcement Learning (PIRL) — systematic comparison of physical-prior integration strategies against standard RL baselines across representative simulated physical domains.
-
Toward Sim-to-Real-Ready 3D Worlds: Gaussian Reconstruction, Object-Centric Segmentation, and Physics Grounding for Reinforcement Learning in Isaac Sim / Isaac Lab
Investigation of real-to-sim and generative-to-sim workflows that combine WorldLabs Marble scene generation with NVIDIA Isaac Sim — bridging the visual/physics gap between Gaussian-splat reconstructions and simulatable robotic environments.
-
Bridging the Sim2Real Gap: Transferring Robot Learning Policies from NVIDIA Isaac Sim to the SO-101 Manipulator
Transfer of robot-learning policies trained in NVIDIA Isaac Sim to the low-cost SO-101 manipulator, with a focus on closing the sim-to-real gap for imitation- and reinforcement-learning-based control.
-
Continual Learning in V-JEPA World Models: Balancing Stability and Plasticity through Self-Modifying Memory
Investigation of continual-learning strategies for V-JEPA self-supervised video world models to mitigate catastrophic forgetting across sequential tasks.
-
Proof of Value robotischer Systeme im Kernkraftwerk Leibstadt: Konzeption und Anwendung eines ganzheitlichen Bewertungsmodells
Holistic assessment framework for evaluating the value proposition of robotic systems in nuclear power plants, applied to a case study at Kernkraftwerk Leibstadt.
-
Detection of Marine Debris Using Synthetic Training Data: A Simulation-Based Approach in NVIDIA Omniverse with Experimental Validation in Real-World Conditions
Automated synthetic-data pipeline in NVIDIA Omniverse for annotated underwater images used to train YOLOv8 marine debris detectors. Production of a 4,500-image synthetic dataset and evaluation of multiple strategies (domain augmentation, data blending) for closing the sim-to-real gap against the real-world DeepPlastic dataset.
-
Evaluating Generalization of CNN Architectures under Domain Shift in Trash Classification
Computer-vision pipeline for automated classification of waste items to support recycling workflows.
-
Theoretische Analyse von KI-Algorithmen zur Punktwolkenklassifizierung im Kontext des Scan-to-BIM-Prozesses
In cooperation with Heberger HTI GmbH. Study of point cloud-based methods for Building Information Modeling (BIM) from indoor 3D scans.
-
Design and evaluation of a cost-effective edge AI node for predictive maintenance in small and medium sized enterprises
Survey of edge-computing architectures and their constraints for on-device inference in autonomous systems with prototype development of an STM32 based anomaly detection application.
-
Reinforcement Learning in speicherbasierten Spielumgebung: Entwicklung und Evaluation eines PPO-Agenten am Beispiel von Pokémon Red
Reinforcement-learning agents trained on Pokémon-style game environments to study exploration strategies and reward shaping.
-
Comparative Analysis of the Sim-to-Real Gap in Different Robotic Simulation Environments
Literature review comparing robot simulators for RL-based manipulation. Conclusion: Isaac Sim leads for perception-heavy tasks and large-scale synthetic data; MuJoCo for contact-rich state-based control.
-
Humanoide soziale Roboter – Wie beeinflusst das Design ihre Akzeptanz bei Menschen?
Survey of social humanoid robot platforms and human–robot interaction design considerations.
-
Enhancing Humanoid Robots Gait Using Unsupervised Reinforcement Learning: Real-time simulation and gait generation of humanoid robot
Modular perception-to-control pipeline for the Fourier Intelligence GR1-T2 humanoid robot (44 DoF) in NVIDIA Isaac Sim. A four-stage architecture translates monocular human video into segment-level targets tracked by an RL policy.
-
Bridging the Sim-to-Real Gap for Monocular 3D Object Detection via Foundation Model-Guided Unsupervised Domain Adaption
Study of unsupervised domain-adaptation methods for deep-learning models operating under distribution shift between source and target data.
-
Evaluation of Existing Deep Learning Methods for Cross-Calibration of LiDAR and Camera Considering a Possible Scan Pattern-Induced Bias
In cooperation with Robert Bosch GmbH. Investigation whether end-to-end deep learning methods for LiDAR-camera extrinsic calibration suffer from shortcut learning due to recognizable LiDAR scan patterns. Findings confirmed by concurrent peer-reviewed publications.
This thesis led directly to a jointly authored journal publication submitted to ICMERR 2026.
-
The Impact of Remote Work on Engineering Team Collaboration and Productivity
Empirical study of how remote work arrangements affect collaboration patterns and productivity in engineering teams.
-
Theoretische Analyse und praktische Implementierung eines Hardware-in-the-Loop-Systems zur 3D-Objekterkennung als Unterstützung des Inbetriebnahmeprozesses anhand einer Pick-and-Place-Anwendung
In cooperation with pester pac automation GmbH. Implementation of a Hardware-in-the-Loop system combining a real stereo-vision camera, an industrial PLC, and a virtual robot simulation for pick-and-place with sub-millimeter X/Y precision.
-
Optimal Sensor Placement in Thermal Compensation of Industrial Robots
Development and comparison of four systematic methods for selecting minimal temperature sensor sets on a Kuka KR30-3 to compensate thermal-induced positioning drift in cooperation with Industrial Robotics Lab Zürich (IRZ) at ETH Zürich.
-
Technische Bewertung und prototypische Umsetzung einer automatisierten Anpassung des Adaptiven Fahrassistenten zur individuellen Optimierung der Fahrweise
In cooperation with CARIAD SE. Personalization of Adaptive Cruise Assist with Machine Learning and empirical validation on real-world fleet data.
-
Study on Autonomous Delivery Robots: Limitations of Pathfinding and Obstacle Detection Algorithms, with a Case Study on Yandex Rover
Survey and case study of last-mile autonomous delivery robots (Starship, Kiwibot, Nuro, Yandex Rover and others). In-depth analysis of Yandex Rover’s LiDAR-based navigation pipeline.
-
Lane Detection and Tracking: Machine Learning for Object Detection and Recognition
Deep-learning approaches for lane detection and tracking in autonomous-vehicle camera pipelines.
-
Using Machine Learning for Ground Navigation Estimation System
Development of a hybrid navigation algorithm combining Extended Kalman Filter sensor fusion (IMU, GPS, odometer) with an LSTM network to maintain position accuracy during GNSS outages.