Clément Callaert
CentraleSupélec, Mathematics and Data Science · M2 Mathematics and AI, Université Paris-Saclay
From January 2027: robot learning, LEMA group, National University of Singapore
I work on generative models of dynamics: probability paths and flow matching, latent world models for planning, and how these behave when the sampling budget is small. I want these methods to end up in general-purpose robots that keep adapting in real settings.
Outside research, I read history, geopolitics, economics and physics; more on the about page.
The projects below are personal work, not lab research. Each has a public repository, a short report, and an explicit note on what the result does not show.
Projects
Few-step field regularity
At a fixed number of function evaluations, lower averaged Jacobian regularity does not reliably imply lower sampling error. The frozen grid has 12 geometry–solver cells, each evaluated at NFE 8, 16, and 32. Five cells contain ranking inversions, spanning 14 of the 36 paired path comparisons. Exact Gaussian systems only.
Sole author. Accepted as a poster at the NeurIPS 2026 GDDL workshop, which is non-archival.
Two more on the research page, with methods and limitations in full.
Notes
Notes on latent world models and planning updated September 2026
Contact
clement [at] clementcallaert [dot] com. I am happy to discuss any of the projects above, including the ones that did not work.
Set in Source Serif 4. Built with Quarto; source. These are personal projects, run on my own machine rather than on cluster compute.