Autre
| Résumé : | Automatic design by optimization is one promising alternative to the manual design of control software for robot swarms. Although this approach offers many benefits, it also entails some challenges. One is to find the correct objective function fitting a specified robot swarm mission. Indeed, these objective functions are not always trivial, even for experts in swarm robotics. Ideally, suiting objective functions could be processed for missions as complex as wanted. Furthermore, even non-expert operators could be able to retrieve relevant swarm behaviour for such complex missions without any prior advanced knowledge. Hence in this master thesis, I present Demonstration-Cho, a method to produce robot control software without providing an explicit objective function. This method allows learning a representation of an objective function from a set of human demonstrations. With the learned objective function, Demonstration-Cho automatically generates control software, using AutoMoDe-Chocolate, leading to desired collective behaviours. In this master thesis, I assess Demonstration-Cho in four missions, providing only demonstrations of the spatial distribution of the swarm at the end of the mission’s time. The results of the experiments show that Demonstration-Cho can design control software with comparable performances to those generated with EvoStick and AutoMoDe-Chocolate with the explicit objective function of the missions. |
