Mobile Robot Cognitive Map Building Based On Episodic Memory

For autonomous mobile robots, mapping, localization and navigation are the three basic but important tasks, and they are hot issues of the research field. This requires the robot using sensors and information input to complete cognitive learning of the environment so as to fulfill further tasks such as target navigation, which is a huge challenge in real natural environment. Besides, in nature, human beings and other animals have a strong cognitive ability to the complex environment, they can learn knowledge of the environment and to complete specific tasks based on memory. Therefore, in this paper, we use the relevant theories of episodic memory explore the related models and methods for the robot's environmental learning.Based on the characteristics and function of the biological cognitive map, considering the activation and episodic memory related to the hippocampal neurons, combining the relevant background of biological environmental cognition and navigation, a model of episodic memory is proposed for mobile robot cognitive map building and meet the needs of the navigation. With in the field of computer vision, image feature extraction and matching technique is used to form the mobile robot scene perception, and based on biological path integration characteristics use competition attractor network to form the robot pose perception, besides, abstract robot behavior patterns, then maps to the state neural to form the episodic memory of the robot.For the cognitive learning problems of the mobile robot in a real environment, a generalized learning algorithm based on episodic memory is put forward, and an episodic memory network is organized to let the robot form a description of the environment in the topology and geometry. First the general process of episodic memory network structure and learning are described, then its detailed function is analyzed, to fulfill the re-localizing, the generation of geometric relationships of state neurons, the process of the dynamic information in the environment and the loop closure of the cognitive. Then, robot navigation is analyzed based on episodic memory, the global path planning of the robot is realized, and a path correcting mechanism is put forward to plan the action of the robot.The robot operating system (ROS) and Turtlebot2 mobile robot are utilized to conduct the experiment of the method above and verify its effectiveness and practicability. First a small loop cognitive learning is conduct, and the function of the algorithm is analyzed based on the result. Then compared the cognitive map and the mature probability between map features, Then the characteristic of the cognitive map and grid map is analyzed and their experiment is conduct under several typical environment. Finally, based on the cognitive map of the environment robot navigation task is completed, and the dynamic information of the environment is processed in some degree.