Visual predictive control for differential drive robots with parallel implementation on GPU
Résumé
This work focuses on the control of a camera mounted on a differential drive robot via a VPC (Visual Predictive Control) scheme. First, an exact model of the visual feature prediction is presented for this robotic system. Next, relying on the equivalent command vector concept, a parallel implementation on a GPU (Graphics Processing Unit) of the computation of the cost function and its gradient is presented. Finally, results show that the proposed approach is more accurate than the ones classically used and can be up to six times faster than CPU-based (Central Processing Unit) one for large prediction horizons and numerous visual features. It then becomes possible to implement a VPC controller running sufficiently fast to perform a navigation tasks, while guaranteeing the closed-loop stability by relying on large prediction horizons.
Domaines
Robotique [cs.RO]Origine | Fichiers produits par l'(les) auteur(s) |
---|