Towards a versatile framework to integrate and control perception processes for autonomous robots
Résumé
Perception is at the heart of autonomous robots, as it is the way through which the decision-making processes get information on the environment and the robot itself. In the last years, the availability for different types of sensors in robotics has greatly improved, allowing the integration of a significant variety of sensors on the same platform. Jointly, the state of the art of data processing and data fusion became much richer, offering a broad choice of solutions. Despite this richness, data fusion and perception processes are still tailored, by the robotic engineers, to the task they are needed for, with the sequences of data processes being defined by initially hardcoded scripts. Yet, perception is an active process: deciding which data to acquire and how to process them yields the possibility to optimize the throughput of perception, that is the relevance and quality of the information provided to the decision making processes. It is indeed short sighted to think that the configuration of the perception processes does not have to be changed regardless of the task a robot is carrying out, or the context within which the task is executed. Even though there have been multiple efforts in active perception, there is the lack of a system abstracting from the type of sensor or from the nature of the task. To the best of our knowledge, there exist no work on defining a generic architecture modeling perception process being sensor and task agnostic. In this paper, we present our ongoing work on the definition of an approach to design an architecture controlling a broad perception layer, that allows to dynamically assemble, sequence and control perception processes on board an autonomous robot.
Domaines
Robotique [cs.RO]
Origine : Fichiers produits par l'(les) auteur(s)
Loading...