Evolving an adaptive controller for a quadruped-robot with dynamically-rearranging neural networks

Abstract

As highly complicated interaction dynamics exist, it is therefore extremely difficult to design controllers for legged robots. Evolutionary robotics is one of the most promising approaches, but there still exist several problems that have to be solved. One of the critical problems is that evolved agents generally tend to over adapt to their given environments through the evolutionary process. In other words, they lack rich adaptability. Therefore, it is highly necessary to establish a method that enables one to efficiently construct adaptive controllers that can cope with different situations. For this purpose we introduce the concept of neuromodulators, allowing the evolvement of neural networks which can adjust not only the synaptic weights, but also the structure of the neural network by blocking and/or activating synapses or neurons. We apply this concept to create an adaptive controller for a quadruped robot which not only can walk forward but also regulate the torque output applied to each joint, according to the current situation.