Liu, Y., Lerch, L., Palmieri, L., Rudenko, A., Koch, S., Ropinski, T., Aiello, M. (2025). Context-Aware Human Behavior Prediction Using Multimodal Large Language Models: Challenges and Insights. IEEE International Conference on Robot and Human Interactive Communication (RO-MAN) 2025.

Project Page: CAP-MLLM - Code: boschresearch/cap-mllm - Paper: arXiv:2504.00839.

Generating realistic tree images with Stable Diffusion and ControlNet

• Description: Trained a custom ControlNet allowing to guide tree image generation with a hand-drawn tree skeleton condition
• Team: Solo
• Technologies: Generative AI, Diffusion Models, Python (PyTorch), HuggingFace

Feedback Networks for Robust Object Detection

• Description: Added feedback connections in Faster R-CNN's ResNet backbone to improve robustness against image corruptions
• Team: Solo
• Technologies: Computer Vision, Convolutional Neural Networks, Python (PyTorch)

Learning to Query Social Media via Interpretable ML

• Description: Combining natural language processing and optimized sparse decision trees to improve social media API queries
• Team: 4
• Technologies: Explainable AI, Decision Trees, Python (scikit-learn)

Monte Carlo Tree Search (MCTS) tictactoe

• Description: Monte Carlo Tree Search (MCTS) Python implementation with tictactoe and Gridworld Gymnasium environments
• Team: 2
• Technologies: Reinforcement Learning, Python

Real-time ML Weather Predictions with Apache Kafka on AWS

• Description: ML predictions on real-time weather data using Apache Kafka on AWS, investigating the impact of resource constraints on the performance of the streaming application pipeline.
• Team: 4
• Technologies: Data Science, MLOps, Kafka, AWS

Statusinfo (Lechvalley)

• Description: Building a web platform on which businesses can transparently communicate the order status to their customers
• Team: 3
• Technologies: Web Development, Frontend (Vue.js), Backend (Laravel), Product Development

TUM.ai x Aleph Alpha: BenchPress Makeathon

• Description: Build a Llama 3.1 based coding agent to solve coding problems evaluated on the APPS dataset
• Team: 4
• Won 1st place in the hackathon

TUM.ai Makeathon 2025: OpenAI Open Track Challenge

• Description: ReplyGremlin - Email handling voice assistant that helps reply to and organise emails, all in a real-time conversational experience.
• Team: 4

EDTH MDS Hack: SE3 Labs Challenge

• Description: Visual drone-map localization pipeline using SuperPoint for keypoint extraction and LightGlue for matching.
• Team: 3

TUM.ai x BKW Engineering Hackathon

• Description: FachPlanner - AI-powered building engineering planning and cost estimation in the construction industry.
• Team: 4