Title:	Asymptotically stable flow matching
DNr:	Berzelius-2025-274
Project Type:	LiU Berzelius
Principal Investigator:	Noémie Jaquier <jaquier@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2025-09-01 – 2026-01-01
Classification:	10210
Keywords:

Abstract

This project aims at developing novel machine learning techniques for learning expressive and stable robot motion policies that can handle high-dimensional, complex, and multi-modal data. To do so, we build on Riemannian flow matching policies, a deep generative model for robot motion generation introduced in our previous work. In this project, we propose to introduce additional inductive bias inspired from control theory and dynamical system to ensure that the learned policies are globally asymptotically stable, allowing safe motion generation with theoretically guarantees. We aim to test our approach on several state-of-the-art datasets to learn end-to-end visuomotor robot policies. We are asking for a continuation of the previous project to finish the experiments that were initially planned but we did not yet managed to run due to delays.