In this project, the use of a reinforcement learning through an hierarchical neural network is proposed to manage behaviors of a robot swarm in order to enhance their collective actions to execute tasks. The chosen reinforcement model is able to balance the influence of previously implemented behaviors through interaction with the environment. Each robot has its own neural network, acquiring its knowledge individually through this interaction and also by sharing information with neighboring robots. In order do evaluate effectiveness, an escorting task is given to be done balancing two behaviors: area coverage through Centroidal Voronoi Tesselations and mainentance of a distance between a robot and a given target. To test the proposed approach, escort scenarios will be simulated in the Player/Stage environment.
News published in Agência FAPESP Newsletter about the scholarship: