Institut des Systèmes Intelligents
et de Robotique


Sorbonne Universite



Tremplin CARNOT Interfaces



THESIS TITLE : The transferability approach: an answer to the problems of reality gap, generalization and adaptation


KEYWORDS : evolutionary robotics , multi-objective optimization algorithms, surrogate models, reality gap problem, generalization , adaptation


30/11/2011 à 13h30
UPMC – Campus Jussieu, 4 Place Jussieu - 75005 Paris
Bâtiment Atrium RdC, Salle de visioconférence 2



The design of controllers for robots that have to deal with unknown or poorly controlled environments is a difficult engineering problem. Evolutionary robotics tackles this challenge by developing automatic methods to design controllers that are based on black-box optimization processes using evolutionary algorithms. In this context, the performance values are estimated for each controller either directly on the robot or with a simulation. Assessing performance with the real robot and in all the possible situations is usually not consistent with the evaluation budgets required by these optimization methods. We here propose a general approach, which combines a multi-objective optimization process in a fixed simulation model, a surrogate model and a few experiments on the robot. This transferability approach can be applied to any optimization process conducted in a simplified environment (simulation) for a targeted full environment (robot). It looks for controllers that maximize two objectives: the performance in simulation and a transferability objective, which reflects how well the behavior observed in simulation matches the behavior on the robot. The second objective is estimated by a surrogate model built by performing a few transfer experiments on the robot while optimizing controllers. The approach is applied on three open problems from evolutionary robotics: the reality gap problem, the optimization of controllers with generalization abilities and the adaptation of a robot to its environment.