18 décembre 2025 à 10h30
Model Predictive Control (MPC) is a powerful framework for robot control thanks to its ability to reason over the future and to update decisions online based on sensory feedback. While its effectiveness in generating complex motions is now well established, MPC remains fundamentally limited in tasks involving controlled physical interaction. The main challenge lies in predicting the evolution of contact forces, which requires dynamic interaction models that are often inaccessible, inaccurate, or unsuitable for real-time optimization. In this talk, I will present a new paradigm that integrates force feedback directly into the MPC loop, enabling robots to regulate contact interactions predictively and safely. I will first revisit classical nonlinear MPC and present the first hardware demonstrations of high-frequency, closed-loop MPC on torque-controlled robots, highlighting why standard formulations inherently fail to incorporate force feedback. I will then introduce new approaches based on state augmentation and online estimation, which overcome this limitation and allow MPC to exploit measured efforts systematically. These contributions include advances in constrained numerical optimal control as well as experimental methods for real-time force estimation. Finally, I will outline my ongoing work on tactile sensing and robust interaction, where I extend these ideas beyond force feedback to address touch-based perception, tactile servoing, and scenario-based MPC for human-robot collaboration. I will conclude by presenting my broader research vision: tactile-aware predictive control as a foundation for safe and adaptive physical interaction with uncertain environments.
Bio : Sébastien Kleff received the M.Sc. degree (“Diplôme d’Ingénieur”) from École Centrale de Nantes and the M.Sc. degree in Electrical Engineering from Shanghai Jiao Tong University in March 2019. He received the Ph.D. degree from New York University in May 2024, where he was co-advised by faculty at NYU and LAAS-CNRS (France). From 2020 to 2022, he conducted part of his doctoral research on-site at LAAS-CNRS. He is currently a postdoctoral researcher at the Inria Centre in Bordeaux. His research interests include model predictive control, physical interaction control, and human-robot collaboration.
Inria
français
English