MASTER THESIS “SCENE-AWARE CONDITIONAL DIFFUSION MODELS FOR ROBOT MOTION PLANNING”

As Austria's largest research and technology organisation for applied research, we are dedicated to make substantial contributions to solving the major challenges of our time, climate change and digitisation. To achieve our goals, we rely on our specific research, development and technology competencies, which are the basis of our commitment to excellence in all areas. With our open culture of innovation and our motivated, international teams, we are working to position AIT as Austria's leading research institution at the highest international level and to make a positive contribution to the economy and society.

Our Center for Vision, Automation & Control located in Vienna invites applications for a master’s thesis. The Center for Vision, Automation & Control exploits the opportunities provided by automation and digitisation to initiate and advance innovation for industry, primarily in Austria and Europe. Our main goal is to take away the monotonous, heavy, difficult and dangerous aspects of people’s work through innovation. We help our partners by making technical systems more robust, reliable, flexible and easy to use. At the same time, we increase the resource efficiency of industrial processes by reducing waste, product failures and emissions.

Our Competence Unit Complex Dynamical Systems focuses on the development and deployment of algorithms to control various types of systems. Starting from low energy applications like electronics and drive systems, via (large-scale) robotics to heavy industrial applications. We are working in projects in cooperation with national and international research organszations and companies leading their market sector.

• Diffusion models are inspired by non-equilibrium thermodynamics. They define a Markov chain of diffusion steps to slowly add random noise to the data, and then learn to reverse the diffusion process to construct desired data samples from the noise. Different from other generative models such as VAE or flow models, diffusion models are learned using a fixed procedure, and the latent variable has high dimensionality.
• Recently, diffusion-based generative models have proven successful in image processing, in reinforcement learning and robotics, e.g., motion planning, task planning, and offline reinforcement learning.
• Two important properties of diffusion models that we would like to explore are their ability to encode multimodal expert trajectories (from optimization-based trajectory planning) and to inherit gradients of multiple costs expressed in optimization-based trajectory planning.
• In this Master thesis, you will work on developing a machine learning algorithm using diffusion models for the trajectory planning task of a manipulator (KUKA iiwa 14 LBR 820) and/or a timber crane considering the surrounding environments such as structures and obstacles.
• You will broaden your knowledge and skills in the field of motion planning, machine learning, and automation & control theory.
• You will train in scientific work and have the opportunity to discuss your experiences with scientific interdisciplinary experts from the field of automation & control, machine learning.
• You may publish your results at a conference or in a journal.

• Ongoing master's studies in the field of Robotics, Mechatronics, AI, or a comparable technical field.
• Good knowledge in Automation & Control, Machine Learning
• Programming experience in Python / C++, Matlab
• Knowledge of Pytorch is advantageous
• Good knowledge of English or German in word and writing

• Duration of the master’s thesis project: 6 months
• EUR 534,-- gross per month for 12 hours/week based on the collective agreement. There will be additional company benefits. As a research institution, we are familiar with the supervision and execution of master theses, and we are looking forward to supporting you accordingly!