Mikel Zhobro

I’m a PhD student with Prof. Georg Martius and since November 2023, I’ve been part of the Autonomous Learning Group at the University of Tübingen and MPI‑IS. Previously, I spent some time at the EV under Jorg Stückler and LDS under Michael Muehlebach. Not so long ago, I earned an M.Sc. in Electrical Engineering from RWTH Aachen and an M.Sc. in Computer Science from KTH Royal Institute of Technology. My main focus is on spatial intelligence. In particular, I am developing temporal world models with inherent 3D spatial consistency both for generation and downstream robotics tasks. In the long term, I aim to scale these models to videos in the wild in a morphology‑agnostic way, from humans to robotic arms and beyond. Outside of my research, I am fascinated by open-ended learning and keep an eye on the latest developments in the field.


News

Publications

Learning 3D-Gaussian Simulators from RGB Videos

Learning 3D-Gaussian Simulators from RGB Videos

Mikel Zhobro, A. René Geist, Georg Martius
ICML, 2026 New!

An action-conditioned 3D world model learned from multi-view RGB videos, coupling a particle-based dynamcis backbone with 3D Gaussian Splatting rendering.

Learning Physics from RGB Videos with scalable Point-Transformers and 3D Gaussian Splatting

Learning Physics from RGB Videos with scalable Point-Transformers and 3D Gaussian Splatting

Mikel Zhobro, A. René Geist, Georg Martius
4DV Workshop @ CVPR, 2025

Precursor to 3DGSim: a non-action-conditioned 3D world model from multi-view RGB videos. First demonstration that scalable point-transformers + 3DGS rendering generalize across scenes.

Learning with 3D rotations, a hitchhiker’s guide to SO(3)

Learning with 3D rotations, a hitchhiker’s guide to SO(3)

ICML, 2024

Guidance on selecting the right rotation representations on deep learning models. The focus is SO3 representation as input or output of a neural network.