Swiss Ai Research Overview Platform
Il existe de plus en plus d’interactions entre biologistes et roboticiens pour étudier et répliquer les impressionnantes capacités de locomotion des animaux. Un aspect important qui n’est toujours pas bien compris chez l’animal est de savoir comment les parties supérieures du cerveau (cortex moteur, cerebellum, …) interagissent avec la moëlle épinière pour planifier et apprendre des mouvements, en particulier pour combiner les capacités d’anticiper le futur et celles de rapidement réagir à des perturbations. Similairement, les roboticiens font face aux difficultés de correctement programmer les robots pour apprendre, planifier et contrôler des mouvements complexes. Le but du projet est de développer des modèles computationnels qui (i) répliquent la réactivité de la moelle épinière avec les capacités d’anticipation et d’apprentissage du cerveau, (ii) permettent de tester des hypothèses par rapport au contrôle du mouvement chez les animaux vertébrés à un niveau conceptuel, et (iii) servent de base pour contrôler, planifier et apprendre des mouvements dans des robots. Ces modèles seront testés dans un robot quadrupède traversant un parcours avec obstacles et trous.
Driven by a desire to understand and replicate the amazing locomotor abilities of animals, there are increasingly useful interactions between biologists and engineers. One important aspect that is still poorly understood in animals is how higher centers such as the motor cortex and the cerebellum interact with spinal cord dynamics to learn and achieve both anticipatory and reactive behaviors. Similarly, roboticists face the challenge of properly implementing learning, planning, and control for complex motor skills. The goal of this project is to develop a computational architecture that (i) merges spinal-cord-like dynamics with higher-level planning and learning, (ii) allows testing hypotheses about animal motor control at a conceptual level, and (iii) can serve as basis for controlling, planning, and learning rich motor skills in robots. The architecture will be evaluated with a scenario that involves control and learning in a quadruped robot crossing a parkour.Using this scenario, the project will contribute to robotics by developing a control architecture that is fast, robust, computationally light, suitable for a distributed implementation, and that can benefit from reflexes and pattern generation offered by spinal cord-like dynamics while at the same time allow learning and planning. The project will also address important scientific questions about animal control such as how the brain performs planning while taking into account spinal cord dynamics and how complex motor skills can be learned and gradually made automatic. Finally the project will contribute to benchmarking by systematically comparing the control architecture against alternative approaches, and by making all software available in open source.
Last updated:18.06.2022