A Robot Navigation Algorithm Based On Cognitive Mechanism Of Rat-brain Hippocampus

Autonomous localization and navigation have always been important research directions in the field of robotics.SLAM,as the mainstream navigation algorithm,still needs a lot of computing resources to achieve.But navigation is a breeze for any kind of creature.Even a nematode of 302 neurons can accurately find food based on its own odor-sensing navigation system.Higher animals evolved into more sophisticated navigation systems.Physiological studies show that the hippocampus is the core region of highermammals brain for spatial cognition,which plays an important role at building cognitive maps and biological navigation.The main work of this paper is to make mathematical modeling and simulation based on the physiological structure and behavior for the hippocampus of research findings.Experimental recover the phenomenon of neuronal firing in the hippocampus,and build a cognitive map to achieve the physical goal-oriented navigation implementation.The main research work includes three aspects:(1)Proposed a more rigorous and more physiologically model of space cells based on RATSLAM.According to the present study,there are mainly four types of navigation-related cells in the hippocampal formation: head-facing cells,stripe cells,grid cells,and location cells.In this paper,the firing characteristics of each cell and the information projection relationship in the brain region are modeled to reappear the firing phenomena and the mechanism of information transmission in the brain regions.The velocity and angular velocity are used as input for the model,the first cell toward the angular velocity encoding.Linear velocity and angular velocity are used as inputs to the model.The head direction cells encoding angular velocity.The head direction cells encode the angular velocity.The output of head direction cells and velocity are the input of stripe cells.Stripe cells use the one-dimensional continuous attractor model to form stripe cell firing field by path integration for the current position.The firing signal of stripe cells simultaneously drives different size grid cells composed of two-dimensional continuous attractor models to form grid cells firing field.The The single firing field of place cells is formed by the firing activities of different grid cells through a Gaussian compettion model.The place cells field and the location of the actual physical world is a one-to-one correspondence.The simulation of model can be good at path integrated,but also simulate the firing characteristics of each space cells.(2)A topological cognitive map is proposed which contains RGB-D image information,bio-firing information and related information between nodes.Unlike RATSLAM,this paper uses a depth camera for visual acquisition to make up for the impact of light on closed-loop point detection.At the same time,each cognitive node stores the color information and the corresponding depth information in the current view.FAST was used to extract the feature points of the image during the closed-loop point detection,and then SURF algorithm is used to describe feature points to generate feature point vectors for rough matching.Finally,3D information was used to eliminate the false matching points.When the closed-loop point is successfully detected,the error of space cells path integration is corrected and the activity of the space cell is reset.(3)Proposed a path planning algorithm based on the cognitive map.Combined with the characteristics of topology map,we propose an improved global path planning algorithm--Dijkstra algorithm.The algorithm is improved by adding weighted summation for all kinds of information of each node and physical distance to the target.So we get a shortest and more feature information global path.Then set the inflection point at each corner on the global path.According to the location of the inflection point,the local path planning finds the node with the most characteristic features around the inflection point as the end point of the local path planning.The image feature points were extracted by FAST,then calculate the translation matrix and rotation matrix between two frames to correct the error.In this way,the global path planning is divided into multiple sections.Local path planning can reduce the accumulated error in continuous navigation.The specific implementation of the local navigation is a relatively mature which is dynamic window method.Dynamic window algorithm can effectively avoid static and dynamic obstacles to reach the target point quickly.