Barate, Renaud (2008) Apprentissage de fonctions visuelles pour un robot mobile par programmation génétique. PhD thesis Algorithmique de Vision, Laboratoire Électronique et Informatique (ENSTA-UEI), Paris VI p.149.

Full text available as: - these-barate.pdf ( 5176 Kb )

Licence: Copyright

Abstract

Existing techniques used to learn artificial vision for mobile robots generally represent an image with a set of visual features that are computed with a hard-coded method. This impairs the system’s adaptability to a changing visual environment. We propose a method to describe and learn vision algorithms globally, from the perceived image to the final decision.

The target application is the obstacle avoidance function, which is necessary for any mobile robot. We formally describe the structure of vision-based obstacle avoidance algorithms with a grammar. Our system uses this grammar and genetic programming techniques to learn controllers adapted to a given visual context automatically.

We use a simulation environment to test this approach and evaluate the performance of the evolved algorithms. We propose several techniques to speed up the evolution and improve the performance and generalization abilities of evolved controllers. In particular, we compare several methods that can be used to guide the evolution and we introduce a new one based on the imitation of a recorded behavior. Next we validate these methods on a mobile robot moving in an indoor environment. Finally, we indicate how this system can be adapted for other visionbased applications and we give some hints for the online adaptation of the robot’s behavior.

Statistiques de consultation

Repository Staff Only: edit this item