A control scheme based on decomposition to primitive behaviors and a reinforcement learning technique
Description
The control of actions for a dynamic non-linear system is a very difficult task to achieve, because is practically impossible to describe in a mathematical model all the physical characteristics and dynamics for the own system, and the characteristics and dynamics of the external environment that affect the performance of such system Autonomous robotic navigation is defined as the task of finding a path along which a robot can move safely from a source point to a destination point in an obstacle filled environment, and at the same time execute the actions to carry out the commands in a real or simulated world In particular, the design of a controller for a mobile robot, navigating in a dynamic environment presents serious challenges to the designer, because of the unstructured and complex characteristics of the environment where the robot is navigating, and the non accurate information about the state of the robot, which is mainly due to the unpredictability of events, noisy environment, delays in the process of sensory information, and the lack of an accurate description of the environment. Thus, obtaining an accurate state of the environment and system based on this information is difficult In this work we propose a self-improving, self-learning navigation system that uses a Reinforcement Learning technique implemented with a Neural Network that is used in conjunction with a modular architecture, that is used to control the fusion and combination of behavioral primitives, in order to execute and control a more complex set of actions performed by the mobile robot. The presented architecture allows the execution of the various behaviors either independently or simultaneously Our main goal is to design a controller for an autonomous mobile robot that uses its own internal motivations such as punishments and rewards to develop cognitive abilities. The robot should find its own solutions to the problems, without any help of external agents guiding its decisions Our approach makes the coupling of actions and perceptions of the outside world as a basis for developing cognitive capacities for the robot. This approach implies that the observed behaviors of the mobile robot should emerge as the effect of the interactions between the robot and the environment on its internal structure. This means that the information and knowledge that system obtains should not be programmed