Research On The Target-oriented Robot Navigation Model Based On The Cognitive Mechanism Of Rat Hippocampus

Rapid and accurately navigated to the predetermined target position(such as homing and foraging) is an important ability to the survival of higher animals. Physiologic studies have shown that the hippocampal formation is the core brain areas for spatial cognitive,which plays an important role in declarative memory and spatial navigation.The study of information processing and cognitive mechanisms involved in spatial cells in rat hippocampal formation provides a road for the development of computational models and robotic spatial cognition of goal-directed navigation, which is current research hotspot and difficulty in the field of artificial intelligence.Based on the research achievements of the behavioral and neurophysiological mechanisms of the rat hippocampal formation,Our research group imitates the cognitive mechanism of hippocampus to study on the methods of goal-oriented navigation model of autonomous mobile robot and its physical embodiment.In this paper, a neural network model based on spatial cells is constructed, which is applied to building the accurate cognitive map. Then we use the cognitive map to study the goal-oriented bionic robot navigation model when the model without the exogenous inputs. The main achievements of this paper are as follows:(1) A path integration method based on discharge mechanisms of spatial cells has been proposed.According to the relationships of information transmission between the four kinds of spatial cells in the hippocampal formation, the mathematical models of each spatial cells has been established, and then a total neural network model of the whole spatial cells has been constructed.The inputs to the neural networks is the speed and direction information, which can generate the band cells’ periodicity firing fields using a one-dimensional(1D) ring attractor model; Then the activities of the band cells in the model project nonspecifically to grid cells, which drives the move of the attractors in the neural sheet to form the grid-cell-like responses to cover the whole environment. Through a process of competitive neural network model, a subset of grid cell population activities will be selected to form one major peak of place cell population activities, which expresses the visited space environment. The simulation results show that neural network model can be used to integrate the self-motion cues to simulate the discharge characteristics of spatial cells.(2) A method which uses visual cues to correct the accumulative errors intrinsically associated with path integration process has been proposed to build accurate cognitive map. The visual cues of environments come from the RGB-D sensor can be used to closed-loop detection. The RGB-D images are used to correct the path integration errors and reset the place cell population activities as well as the associated grid cell population activies to previous locations and directions when loop closures are detected. The robot’s spatial coordinates are calculated from the place cell population activities and the cognitive map are constructed by associating the major peak of the place cell population activities with corresponding visual cues and its topological link. Physics experiments show that this method can build accurate cognitive map in a wide range of space.(3) A robot’s goal-directed navigation model which can use the direct reinforcement learning algorithm to navigate in continuous state and action space has been proposed. Based on the rat’s Morris water maze experiment results, we research the rat’s goal oriented navigation mechanism with the rat can get no exogenous inputs from the external environment. For the problem of the goal-directed navigation in continuous state and action space, a spiking neural network model from hippocampal place cells to the putative action cells in prefrontal cortex is proposed based on the firing characteristics of place cells and the information cycles between the hippocampus and medial prefrontal cortex. The continuous state and action space are respectively characterized by a population of place cells and action cells, and the direct reinforcement learning algorithm combined with the spike response model has been used to goal-directed autonomous navigation. The simulation results in Morris watermaze task show the algorithm used in the model can solve the problem of the goaldirected navigation in continuous state and action space and obtain a better performance when compared to the classic methods. When we change the number of the action cells, the convergence properties of the model are invariant. The model can still navigate to the goal location, when the scale of the watermaze and the goal location have been changed.