MS Thesis in collaboration with DENSO Automotive

Physical AI with Vision Language Action (VLA) Models in Robotics

28.09.2026, Diplomarbeiten, Bachelor- und Masterarbeiten

Modern VLA models perform well on short‑horizon robotic tasks but often struggle with long‑term planning and multi‑step control. A promising direction is to enhance VLA architectures by integrating them with Physical AI models capable of understanding physical effects, e.g. for challenging, unstructured environments.

For such VLA-based physical AI systems, the project aims to evaluate and quantify a robot’s action‑generation capability of a physical AI VLA model for more challenging, unstructured environments, and to develop a clearer understanding how these contribute to more stable and capable long‑horizon action generation.

Research Areas

• Survey state-of-the-art VLA models and world-model approaches
• Implement and benchmark physical variants using a VLA framework
• Develop simulation-based setups for model evaluation in challenging scenarios
• Explore and validate improved integration strategies based on the findings (if feasible)
• Identify and design task-specific fine-tuning strategies to improve model performance (if feasible)

Prerequisites:

• Basic understanding of and strong interest in robotics
• Basic understanding of computer vision
• Ability to understand and run modern deep-learning and robotics codebases
• Proficiency in Python or C/C++
• Experience with physical AI, world models, and simulation is a plus

Kontakt: christian.prehofer@tum.de