Description du projet
Mouvement robotique: concilier la théorie et la pratique
Imaginez un monde où les robots aident les humains accomplissant des tâches telles que courir sur des terrains rocailleux, soulever un canapé et tendre les bras pour récupérer des objets ou encore utiliser un tournevis en se tenant en équilibre sur une échelle. L’un des principaux obstacles à la réalisation de cet objectif est l’absence d’un cadre théorique cohérent pour la locomotion et la manipulation des robots. Le projet CONT-ACT, financé par le Conseil européen de la recherche, vise à élaborer une approche exhaustive qui place les interactions de contact et l’utilisation des informations sensorielles au cœur de la génération et du contrôle des mouvements des robots. En développant une architecture reposant sur des contrôleurs prédictifs en temps réel qui exploitent pleinement les interactions de contact et en analysant les informations sensorielles pendant ces interactions, le projet ambitionne de faire avancer la théorie générale du mouvement des robots qui permet une amélioration constante des performances.
Objectif
What are the algorithmic principles that would allow a robot to run through a rocky terrain, lift a couch while reaching for an object that rolled under it or manipulate a screwdriver while balancing on top of a ladder? Answering this seemingly naïve question resorts to understanding the fundamental principles for robot locomotion and manipulation, which is very challenging. However, it is a necessary step towards ubiquitous robots capable of helping humans in an uncountable number of tasks. The fundamental aspect of both locomotion and manipulation is that the dynamic interaction of the robot with its environment through the creation of physical contacts is at the heart of the tasks. The planning of such interactions in a general manner is an unsolved problem. Moreover, it is not clear how sensory information (e.g. tactile and force sensors) can be included to improve the robustness of robot behaviors. Most of the time, it is simply discarded. CONT-ACT has the ambition to develop a consistent theoretical framework for motion generation and control where contact interaction is at the core of the approach and an efficient use of sensory information drives the development of high performance, adaptive and robust planning and control methods. CONT-ACT develops an architecture based on real-time predictive controllers that fully exploit contact interactions. In addition, the structure of sensory information during contact interactions is experimentally analyzed to create sensor representations adapted for control. It is then possible to learn predictive models in sensor space that are used to create very reactive controllers. The robot constantly improves its performance as it learns better sensory models. It is a step towards a general theory for robot movement that can be used to control any robot with legs and arms for both manipulation and locomotion tasks and that allows robots to constantly improve their performances as they experience the world.
Champ scientifique
- natural sciencescomputer and information sciencesartificial intelligencemachine learningunsupervised learning
- natural sciencescomputer and information sciencesartificial intelligencemachine learningreinforcement learning
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringsensors
Programme(s)
Thème(s)
Régime de financement
ERC-STG - Starting GrantInstitution d’accueil
80539 Munchen
Allemagne