Essential Dynamical Structure in Learnable Autonomous Robots

This paper studies the essential dynamical structure that arises in two different classes of learning of the sensory-based navigation, namely skill-based learning and model-based learning. In skill-based learning a robot learns navigational skills for a fixed navigational task such as homing, while in model-based learning a robot learns a model of the environment, then conducts planning on the model to reach an arbitrary goal. We formulated that the former is achieved by learning the state-action map, and the latter does by learning the forward model of the environment, using recurrent neural learning scheme. The analysis of the dynamical structure from the coupling of the internal neural dynamics and the environment showed that generation of the global attractor is crucial for both learning cases. Experiments were conducted using a mobile robot with a laser range sensor, which verified our assertions in a simple obstacle environment.