Jonathan Sadeghi
I currently work at Odyssey, an AI lab where we’re building interactive video models. Previously, I was a Senior Research Engineer at Bosch (formerly FiveAI). Before that, I was a PhD Student at the University of Liverpool, supervised by Marco de Angelis and Edoardo Patelli. As an undergraduate I read Theoretical Physics at the University of Manchester.
Research focus
My work spans learning algorithms for generative world models and methods for evaluating and improving complex systems. This includes work on perception and generative models, sometimes in embodied settings such as autonomous vehicles. I am particularly interested in reliability, efficient testing, and calibration under uncertainty.
Featured work
Odyssey-2 Max
Odyssey's largest general-purpose world model to date, scaling real-time interactive video generation for more accurate physics and longer, more stable simulations.
CaliBench: Are the Stochastic Dynamics of Video World Models Physically Calibrated?
A benchmark that tests whether video world models reproduce known physical outcome distributions, separating calibration from the ability to generate scoreable videos.
A Step Towards Efficient Evaluation of Complex Perception Tasks in Simulation
A surrogate-based method that makes large-scale testing of complex perception pipelines practical without repeatedly running their most expensive components.
Efficient Training of Interval Neural Networks for Imprecise Training Data
A computationally feasible way to train neural networks with interval predictions, including imprecise training data and statistical reliability guarantees.