Visual and motor learning using a chaotic recurrent neural network: application to the control of a mobile robot

摘要

We give here new results on the possible use of dynamicalrnneural systems in order to produce adaptive behavior, in therncase of a mobile robot interacting with its environment. Arnrecurrent neural network with spontaneous chaotic behaviorrnis the core of our system. With a hebbian covariance learningrnrule, such a network can associate a specific cyclic dynamicsrnto a given static or periodic stimulation. The network is thenrnconnected to visual and motor interfaces of a mobile robot.rnLearning is performed while the robot observes its environmentrnand produces movements. After learning, the systemrnis able to dynamically modify its behavior, depending on therncoherency between its own learned representations and thernexternal inputs. A good matching corresponds to a regularrndynamics and the production of a sequential movement. Arnbad matching produces more irregular dynamics which canrnbe seen as exploratory behaviors; eventually, the dynamicsrngoes back to the learned periodic motor sequence if the robotrnfinds visual stimuli that it has previously learned.