I2SL: Learn How to Swarm Autonomous Quadrotors Using Iterative Imitation Supervised Learning
Résumé
In this paper, a decentralized controller for a quadrotor swarm is presented following the leader-follower principle. The quadrotors embedding the decentralized controller follow a remotely controlled leader. The controller, governing the behavior of a set of followers is learned using an Iterative Imitation supervised learning approach. The novelty of this approach is to build complex policies supporting the flocking behavior for a set of quadrotors while requiring only COTS (Commercial Off The Shelf) wireless sensors. In the first iteration, a set of trajectories is generated using the well-known Reynolds flocking model (adapted by Schilling et al, 2018, to add a migration term); the logs are exploited to enable the follower quadrotor controller to achieve the migration function. In the further iterations, the learned controller is exploited in combination with the Reynolds model; the logs generated are then exploited to learn a follower quadrotor controller achieving both the migration and the flocking functions, as robust as the Reynolds model. The validation of the approach using a Software In The Loop (SITL) environment relying on the Gazebo simulator, confirms that the learned controller enables the followers to accurately follow the leader while collectively satisfying the swarm properties.
Domaines
Intelligence artificielle [cs.AI]
Origine : Fichiers produits par l'(les) auteur(s)