Reverse-engineering the cerebellum for adaptive robot control

This research project proposes modeling, implementing and studying a functional reduced version of the cerebellum. This model, although reduced, must be sufficiently detailed to include different types of cells and functional elements with their characteristic dynamics in order to obtain a realistic model exploiting biological knowledge and computing capacity currently available. Each type of neuron shows specific morphology and membrane properties, which confer different information-processing capabilities. This biologically-plausible cerebellar model will include detailed neural dynamics and properties, such as plasticity and activity resonance. As a result, the impact of specific-neuron-and-network characteristics on cerebellar information processing can be assessed.

The cerebellar model will be integrated into a control loop, which will provide a robotic arm with the torque values required to follow predefined trajectories. The cerebellar model must be able to learn to generate output signals which improve the robot trajectory. This cerebellar model will thus provide an adaptive and biologically-plausible control mechanism which could be suitable for future robotics based on high-inertia actuators and changing conditions. By means of this set-up we will evaluate the model efficacy of learning, memory and pattern recognition features.

This research project is supported by the University of Granada.