Integration of paired spiking cerebellar models for voluntary movement adaptation in a closed-loop neuro-robotic experiment. A simulation study

摘要

Motor control is a very important feature in the human brain to achieve optimal performance in motor tasks. The biological basis of this feature can be better understood by emulating the cerebellar mechanisms of learning. The cerebellum plays a key role in implementing fine motor control, since it extracts the information about movements from sensory-motor signals, stores it by means of internal models and uses them to adapt to the environment. The hypothesis is that different internal models could work both independently and dependently. So far, there have been a few studies that aimed to prove their dependency; however, this hypothesis has not been widely used in robot control. The purpose of this work is to build paired spiking cerebellar models and to incorporate them into a biologically plausible composite robotic control architecture for movement adaptation. This is achieved by combining feedback error learning and cerebellar internal models theories. Thus the control architecture is composed of cerebellar feed-forward and recurrent loops for torque-based control of a robot. The spiking cerebellar models are able to correct and improve the performance of the two-degrees of freedom robot module Fable by providing both adaptive torque corrections and sensory corrections to the reference generated by the trajectory planner. Simulations are carried out in the Neurorobotics platform of the Human Brain Project. Results show that the contribution provided by cerebellar learning leads to an optimization of the performance with errors being reduced by 30% compared with the case where the cerebellar contribution is not applied.