Benchmarking cerebellar control

Abstract Cerebellar models have long been advocated as viable models for robot dynamics control. Building on an increasing insight in and knowledge of the biological cerebellum, many models have been greatly refined, of which some computational models have emerged with useful properties with respect to robot dynamics control. Looking at the application side, however, there is a totally different picture. Not only is there a robot on the market which uses anything remotely connected with cerebellar control, but also in research labs most testbeds for cerebellar models are restricted to toy problems. Such applications hardly ever exceed the complexity of a 2 DoF simulated robot arm; a task which is hardly representative for the field of robotics, or relates to realistic applications. In order to bring the amalgamation of the two fields forward, we advocate the use of a set of robotics benchmarks, on which existing and new computational cerebellar models can be comparatively tested. It is clear that the traditional approach to solve robotics dynamics loses ground with the advancing complexity of robotic structures; there is a desire for adaptive methods which can compete as traditional control methods do for traditional robots. In this paper we try to lay down the successes and problems in the fields of cerebellar modelling as well as robot dynamics control. By analysing the common ground, a set of benchmarks are suggested which may serve as typical robot applications for cerebellar models.