Cerebellar-inspired bi-hemispheric neural network for adaptive control of an unstable robot

In this paper, a cerebellar-inspired adaptive motor controller is constructed, and applied for adaptive control of a two-wheel balancing robot as an example. The controller comprises a feedback proportional and derivative (PD) controller and a realistic bi-hemispheric cerebellar neural network. The cerebellar network was configured based upon current anatomical and physiological knowledge of the cerebellar cortex, consisting of 1560 granular cells (Gr), 10 Golgi cells (Go), 10 basket/stellate cells (Ba/St), and two Purkinje cells (Pk). The network connectivity follows realistic synaptic converge and divergence ratios as close as possible within the limitation in the number of neuron models for real time execution. Each cell is described by a typical artificial neuron model whose output is a weighted sum of the inputs after a sigmoidal nonlinear transformation. The PD controller represents the non-cerebellar component working in tandem with the cerebellum in the brain. In the proposed controller, it provides the error signal for the cerebellar neural network to induce synaptic plasticity at the Gr-Pk synapses as in the real cerebellum. We demonstrate that the proposed cerebellar-inspired controller not only successfully control the balancing robot but also compensates for abrupt asymmetrical perturbations, which neither a PD controller alone nor a cerebellar network with a single hemispheric structure can cope with.