Mihailo Radović

I'm an MSc student in Electrical Engineering and Computing (module: Applied Mathematics) at the University of Belgrade, School of Electrical Engineering. I also work as a Data Engineer at Nordeus.

My research focuses on generative models, such as diffusion and flow matching, and their applications to robot learning. I'm also interested in world models, particularly JEPA-style predictive architectures. I also keep a blog, where I write about what I'm learning and share notes from my research.

I completed my BSc in Computer Science at Union University, School of Computing, in 2025. My bachelor's thesis, supervised by Prof. Nemanja Ilić, focused on flow matching for pixel-art character generation.

Email  /  CV  /  Scholar  /  GitHub  /  LinkedIn  /  Twitter  /  Blog

PhD search: I'm currently looking for a PhD position in generative models, world models, and robot learning. My long-term goal is to help build general-purpose autonomous agents that understand the physical world, learn from experience, and act in it safely, and I'm open to research directions that move toward that goal. If you think I'd be a good fit for your group, I'd love to hear from you by email.

Research

I'm interested in generative modeling (diffusion and flow matching), world models (JEPA), and robot learning. Most of my work is on flow matching: using continuous-time generative models both to generate images and to learn expressive, multi-modal policies for continuous control. Representative papers are highlighted.

Modeling a behavioral cloning policy as a continuous-time flow captures the multi-modal behavior in diverse offline datasets, matching or outperforming Gaussian and diffusion policies on standard continuous-control benchmarks.

A flow matching model with a time-conditioned U-Net that transforms Gaussian noise into 128×128 RGBA pixel-art characters.

Miscellanea

Built on Jon Barron's website template.