Robots that learn to move: why, when, and how?

Jean-Baptiste Mouret

Inria Larsen Team, France

Abstract

The recent advances in deep learning are generating an impressive interest in machine learning, but their influence on robotics is not as strong as we could think (yet), mostly because they require too much data to run online with real machines. In this talk, I will exhibit some situations for which robots can benefit from learning and what constraint robots impose on learning algorithms. Focusing on trial-and-error learning, I will then introduce the work of our team on data-efficient policy search, in particular to allow legged robots to recover from mechanical damage in a few minutes. I will also highlight our preliminary work on combining data-efficient learning with whole-body control.

Biography

Dr. Jean-Baptiste Mouret is a senior researcher ("Directeur de recherche") at Inria, the French research institute dedicated to computer science and mathematics. He is currently the principal investigator of an ERC grant (ResiBots – Robots with animal-like resilience, 2015-2020). From 2009 to 2015, he was an assistant professor ("maître de conférences") at the Pierre and Marie Curie University (Paris, France). Overall, J.-B. Mouret conducts researches that intertwine machine learning and evolutionary computation to make robots that can adapt in a few minutes. His work was recently featured on the cover of Nature ("Robots that adapt like animals", Cully et al., 2015) and it received several national and international scientific awards, including the "Prix La Recherche 2016" and the "Distinguished Young Investigator in Artificial Life 2017".