Reasoning about Perception Uncertainty in Nonlinear Model Predictive Control

In robotics, nonlinear Model Predictive Control (MPC) has emerged as a promising tool to generate complex motions while enabling online adaptation of the robot behavior as the environment changes. However, the lack of efficient computational methods hindered its widespread deployment on real hardware. In practice, MPC formulations and computational methods are often simplified to obtain real-time capable controllers. In this talk, I will present efficient numerical methods to fully leverage the promises of MPC on robots by ensuring safe and globally optimal plans while being aware of the uncertainty resulting from the partial sensing of the environment.

 

Armand Jordana received his M.Sc. degree in Mathematics, Computer Vision, and Machine Learning from Ecole Normale Supérieure Paris-Saclay in 2019. Since Fall 2020, he has been pursuing the Doctor of Philosophy degree in Electrical Engineering at New York University, Tandon School of Engineering, under the supervision of Ludovic Righetti and Justin Carpentier. His research interests include model predictive control, risk-sensitive control, estimation, and robotics.