| Resumo: | There is a growing interest in the development of service and assistive robot technologies for application in domestic and urban environments. Among the required abilities are autonomous navigation and safety maintenance. Machine Learning provides a set of computational tools that have proved useful for robot navigation, such as neural networks, reinforcement learning and, more recently, end-to-end deep learning. This dissertation aims to investigate the problem of mobile robot navigation in a maze-like environment using a reinforcement learning framework. In particular, the work focuses on how to scale reinforcement learning, and Q-learning in particular, to a real-world problem using a physical robot. First, in order to avoid large state-action spaces and long horizons, the robot system is trained using a hierarchical approach in which low-level components (sub-tasks) are sequenced at a higher-level. Second, a dense reward function is designed for robot navigation in a corridor and moving around a corner, providing the robot with more information (prior knowledge) after each action. The experiments conducted, using a simulated and a real robot, show the feasibility of the hierarchical approach in reducing the complexity of the learning task and the role of the reward function in goal specification. Finally, the study provides detailed evaluation about transferring experience in simulation to the physical robot.
|
|---|