Inverse Reinforcement Learning In Distributed Mobile Robotic Architecture

Autonomous mobile robot is one of the most active research topics in artificial intel-ligence. In order to make it possible for autonomous navigation in complex environments, the architecture used in it combines modules of perception, fusion, modeling, planning, rea-soning, and behavior. As technology advances, the architecture of autonomous is constantly improving. During these years, distributed system has become an attractive research domain owing to its fabulous flexibility and high robustness. Due to reasons mentioned above, we concentrate on the distributed system architecture and implement it in an autonomous ve-hicle.This thesis first surveys technological development of mobile robot, and then analyzes and discusses architectures used in it. Based on that, we design and realize a distributed mobile robotic architecture, which models function modules as agents, and classifies them into three groups:perception module, decision module and behavior module. All of the agents are distributed on different computers, and transfers data/message through a platform which supplies message delivery service. Each agent works independently, and all of them are organized as an asynchronous navigation pipeline according to the specified task so as to make system running efficiently. Practical results in a wide range of environments verify that the distributed architecture is flexible and robust.We second introduce reinforcement learning and inverse reinforcement learning those are belong to machine learning algorithms and implement a learning framework for pol-icy generation in the distributed system architecture mentioned above. It applies inverse reinforcement learning to learn the reward functions. In this framework, a set of agents are assigned with different tasks, such as sampling teaching, state decision, feature extraction and reward evaluation. The reward obtained will be exploited by reinforcement learning to achieve optimal policy. Simulated experiments show that this architecture can gain decent policy. In addition, there are many factors that affect the state of environment in complex MDP tasks, therefore it is not easy to select features to represent the state, and furthermore the features are usually in high dimension. Thus it is common to set features manually in traditional inverse reinforcement learning. This thesis achieves automatic feature selection with a dimension reduction method and attempts to attack feature selection problems in a new direction under the inverse reinforcement learning framework.