PhD Student Reasoning Methods for Autonomous Systems

The Fraunhofer Society is one of the world's leading organisations for application-oriented research. 75 institutes develop groundbreaking technologies for our economy and society – more precisely: 32 000 people from technology, science, administration and IT. You know, who comes to Fraunhofer wants and can change something. For yourself, for us and the markets of today and tomorrow. As part of the ''Center for Applied Artificial Intelligence Duisburg'' (short: ZaKI.D), we want to support the regional economy in the transformation in Duisburg and adjacent regions with promising AI technologies. Together with the University of Duisburg-Essen and the KROHNE Messtechnik GmbH, we will help small and medium-sized enterprises (SMEs) from the Fraunhofer Institute for Microelectronic Circuits and Systems to integrate artificial intelligence into their businesses and products. The technological focus of the ZaKI.D is embedded artificial intelligence on resource-restricted devices such as sensors, smart devices or robots. We want to make these safer and smarter, optimize manufacturing processes, increase the quality of products and at the same time minimize the ecological footprint. Generative Artificial Intelligence (AI) is a part of the AI, in which both the conclusions of user inputs and the generation of data play an essential role. When interacting with the physical world, the improvement of reasoning methods for subdividing a user input into small subtasks is a central component to improve the algorithms. With regard to robotic systems, these reasoning methods are used to subdivide the work steps entered by the user for a cobot in such a way that it can execute them step-by-step. The aim is always to execute these algorithms on-premise - at this point you will apply your three-year doctorate.

First of all, you will explore the research stand and develop the theoretical foundations of reasoning in generative AI methods. You design and implement advanced training and fine tuning methods (Reinforcement Learning from Human Feedback, Reinforcement Finetuning, Mixture of Experts, etc.) to improve the reasoning capabilities of LLMs. Based on existing libraries and frameworks (Multimodal: Flamingo / MERLOT / ALPRO / Pipelines for Reasoning in HuggingFace Transformers), you will establish a test and evaluation environment for popular reasoning processes. Using concrete scenarios in household robotics, develop, implement and evaluate new methods in the field of reasoning for robotics applications (Chain-of-Thought (CoT) & Reasoning-Guided generation). You will publish the results of your experimental work as scientific publications and present them at international conferences and in the project meetings with external partners. In close cooperation with our experts, you implement the methods in a generative AI framework that can be used by industrial customers. In addition, you will supervise student and scientific assistants as well as bachelor and master thesis.

Minimum qualification: Highly completed studies (M.Sc. or Uni-Diplom) in computer science, mathematics, physics, electrical engineering or a comparable technical/nature science course Knowledge in the programming language Python and/or C++ and with relevant tools/frameworks such as ROS, PyTorch, Tensorflow, Scikit-Learn or the like. Knowledge of how to use bibilotheques for data analysis, such as Pandas or Seaborn A systematic and independent way of working, coupled with creativity, hands-on mentality and communication security Very good knowledge of German and English for cooperation with our project partners and in an international environment First practical experience with Multimodal Large Language Models (MLLMs) Required qualifications: Master or Diploma thesis in the field of Multimodal Large Language Models (MLLMs) and/or reasoning models. First experiences on embedded AI (conversion with Jetson Orin or AGX) Ideal experience in handling HPC clusters

You work on pioneering topics in a redesigned team with interdisciplinary scientists in the field of machine learning and generative AI In addition to the support provided by an institute-internal specialist, a regular exchange takes place on the scientific state of your work within the framework of a doctorate/vätergremium. During the promotion period, you will be supported by accompanying offers. You will learn to safely apply professional methods of project management and to support the project acquisition. If the dissertation is submitted in due time, the option is to further deepen the research work within the framework of a connection agreement or to switch to other areas. The full-time post as a doctoral student with half-remuneration provides 50% of the time for your doctoral degree and 50% for cooperation in research projects as a scientific assistant. Flexible working hours (running time with integrated core working time from 9:30 - 15:00, Friday from 9:30 - 13:00) and mobile work on up to two days a week for better reconciliation of work and private life Occupational pension (VBL) and grant to the Deutschland-Ticket Job Very good transport with public transport/car and free parking and lockable bicycle parking for employees Family and Occupational Compatibility Support: Mit-Kind-Office, Kindernotpflege und Beratungs zu Homecare-Eldercare etc. in cooperation with the pme family service Corporate Benefits: Benefits Offers from renowned manufacturers and brands