Learning control

To control complex robotic systems, a very convenient solution is often to learn controllers instead of entirely defining them via modelling and control theory or expert knowledge. 
The learning algorithm can typically acquire data from user demonstrations or on its own in the context of adaptive control or reinforcement learning. 
Learning from Demonstration (LfD) has many applications as it provides a user-friendly way to program new behaviors into a robotic system. We have been interested in designing learning algorithms that can ensure that the learned controllers always verify some properties, in particular asymptotic stability. As this property is global and not closed under most basic operations such as addition, it cannot be handled by classical learning algorithms. Solutions exist in the literature, but we proposed a new approach based on efficient constructions of diffeomorphisms to tackle this issue. 
This approach is presented in the following article:

• Perrin, N. and Schlehuber-Caissier, P. (2016), Fast diffeomorphic matching to learn globally asymptotically stable nonlinear dynamical systems, Systems & Control Letters, Vol. 96, pp. 51-59