A simple mechanical system for studying adaptive oscillatory neural networks

Central pattern generators (CPG) are oscillatory systems that are responsible for generating rhythmic patterns at the origin of many biological activities such as for example locomotion or digestion. These systems are generally modelled as recurrent neural networks whose parameters are tuned so that the network oscillates in a suitable way, this tuning being a non trivial task. It also appears that the link with the physical body that these oscillatory entities control has a fundamental importance, and it seems that most bodies used for experimental validation in the literature (walking robots, lamprey model, etc.) might be too complex to study. In this paper, we use a comparatively simple mechanical system, the nonholonomic vehicle referred to as the Roller-Racer, as a means towards testing different learning strategies for an recurrent neural network-based (RNN) controller/guidance system. After a brief description of the Roller-Racer, we present as a preliminary study an RNN-based feed-forward controller whose parameters are obtained through the well-known teacher forcing learning algorithm, extended to learn signals with a continuous component.