Comparative study of affordance-based navigation and model-based navigation: Experimental analysis of learning ability of mobile robot that taps objects with a stick for navigation

Navigation is an important topic in robotics, and various approaches have been proposed thus far. Conventional methods are of two types: those that create and utilize a model of the environment, and those that employ the environment itself instead of a model. For autonomous robots that have learning ability, the difference between these approaches has a significant impact because the learning target differs fundamentally. In the former case, the learning target is to create a model of the environment. Hence, robots learn the positions of obstacles, walls, and so on. In the latter case, the learning target is to explore how the various properties of the environment can be used for navigation. In this paper, we discuss the difference between these two types of learning-based navigation. We employ reinforcement learning as a learning algorithm and conduct simulations and experiments. Consequently, we confirm that the latter mode exhibits a better performance, and in particular, we demonstrate the real-time learning and applicability of the obtained policy for different environments.