学習とロボット 力学系に基づく自律移動ロボットの行動学習

Abstract

This paper describes an experimental results based on our prior-proposed scheme : learning of sensory-based, goal-directed navigation. The emphasis is that learning a task-based behavior could be formulated as an embedding problem of dynamical systems : desired trajectories in a task space should be embedded into an adequate sensory-based internal state space so that an unique mapping from the internal state space to the motor command could be established. In this study, two types of neural architecture were tested for constructing such an adequate internal state space based on sensory information. A simple homing task was conducted on Feed Forward Neural Network that is fed with a regression of sensory sequence as time-delayed manner. A task of complex cyclic routing was conducted on Recurrent Neural Network (RNN) . In both cases, sufficient supervised training of robots generated rigid internal structure of attractor dynamics which realized desired navigations in robust manner, even against limited environmental changes as well as miscellaneous noises in the real world. It was confirmed that RNN architecture has more flexiblity adapting to various tasks.