A self-learning crawling robot

Abstract

Technology is advancing at a fast pace and humanity is becoming more dependent on technology. The increase of computational power in integrated circuit (IC) made it possible for the implementation of more complex algorithms in embedded systems. An area which is constantly improving is Artificial intelligence (Al). Scientists design machines with the ability to behave, learn and act like human beings and have the ability to solve complex problems that involve intelligence, such as pattern recognition in voice, speech or image systems. The aim for this thesis was to implement a crawling robot, which, after interacting with the environment; the robot learns how to move backward or forward without any form of external guidance. The Q-learning algorithm, which is a free model based on reinforcement learning, was implemented on the robotic system so that the robot could learn and recognize movement patterns. This was achieved after a sequence of trials performed by the robot. The results obtained from the various sensors installed on the robotic system, which also included ultrasonic and quadrature encoder, determined the immediate rewards after performing an action, identifying the effects when taking a particular decision. These results were then used by the Q-Learning algorithm to calculate the optimal policy (best path). The main aim of the robot was to achieve its desired goal by maximizing the rewards over time. The sensors together with the algorithm were implemented on a microcontroller. The values obtained from the sensors were used to calculate the Q-Learning algorithm in order to achieve the best path for the robot to move backward and forward.