Research Of Bionic Navigation For Mobile Robots Inspired By Spatial Cognition Mechanism Of Localization Cells

Mammals in nature have a strong ability to recognize and adapt to their environments,they can achieve autonomous localization and navigation by relying on the brain’s intrinsic spatial positioning system.Therefore,brain-like bionic navigation method provides a new technological approach for mobile robot to achieve precise positioning and fast navigation.However,current bionic navigation methods focus on the understanding of brain function,and propose the biological models and test their plausibility.Most of them have insufficient abilities of experience knowledge learning and bionic behavior planning,and they are not well applied to robotic navigation tasks in real environments.Based on spatial cognition mechanism of grid cells and place cells in animal’s brain,we focus on the researches of generalized environment modeling,cognitive map building,navigation behavior planning and other key issues,and which form a brain-like bionic navigation method for mobile robot.The main research contents are as follows:Inspired by the biological basis of episodic memory and the cognitive mechanism of localization cells,the state neurons are abstracted to map high-dimentional scene perception and simulate the organization process of episodic memory,and an episodic memory model integrating the dynamic process of neurons learning and memory formation is established.We encapsulate scene perception,state neurons,pose perception and phase perception to represent experience knowledge,and the incremental events sequence are created to store these experience knowledge,which can realize the generalized modeling of the environment,and effectively overcome the ambiguity problem of environmental perception.The neural activity mechanism of grid cells and place cells are modeled,and the relationship between the firing patterns of grid cells with the position of mobile robot is established.The oscillation phase is proposed to encode the spatial position of mobile robot,and provide accurate pose perception for event.At last,an episodic memory based self-organization learning(EM-SOL)framework is proposed,which can guide the mobile robot to achieve cognitive map building and navigation behavior planning in the natural environments.To improve the adaptability of cognitive map to the change of physical environment,a cognitive map building method based on episodic memory network incremental learning is studied.Based on the neuronal synaptic plasticity and sequence properties,a self-organization learning model is built,and the activation and attenuation characteristics of state neurons are proposed to construct the episodic memory network,which can realize the incremental learning of robotic experience knowledge,and learn the new environments without overwriting the original memory.In this paper,the episodic memory network is abstracted to express the cognitive map,and the incremental cognitive map by self-organization learning is built,which can realize the accumulation,integration and updation of the cognitive map to the change of physical environments.For the accumulative deviation during the long-time motion of mobile robot,a phase reset mechanism is proposed to correct the deviation,which can improve the accuracy of cognitive map.This method shows high map building accuracy and efficiency,it has the advantages of simple parameters adjustment process and wide application range.For the non-specific navigation tasks in the natural environments,the robotic navigation behavior planning method based on episodic cognitive map is studied.Firstly a global path planning strategy based on experience memory is proposed,the collaborative mechanism of context state neurons is introduced to solve the fuzzy problem of state neurons sequence planning.Through events sequence reorganization and optimization,a shortest global path is quickly planned for mobile robot.Then a hierarchical cognitive navigation behavior planning method integrating with head direction cells and grid cells is proposed.The target navigation task is divided into several sub-target navigation tasks,with the coorperation between activity of head direction cells network and similarity of grid cells firing patterns,the mobile robot can reach these sub-targets with high confidence level.Finally a multi-sensor based behavior reasoning strategy is proposed,with which the robot can avoid the obstacles and return to the planned path,making up for the lack of expression of environmental uncertainty in cognitive map.As an extension of research on navigation behavior planning with cognitive map,the mapless behavior planning method based on deep reinforcement learning is also studied.We introduce the animal’s target-oriented navigation knowledge,and propose a heuristic knowledge driven deep deterministic policy gradient algorithm HK-DDPG,which can guide the navigation behavior planning for mobile robot.Aiming at the problems of low utilization rate of sample data and poor training stability of original DDPG algorithm,we introduce the temporal difference error,and assign the different weights to different samples in experience replay pool,a weight based optimal sampling mechanism is proposed.The experimental results show that HK-DDPG algorithm has the characteristics of fast convergence,high learning efficiency and stability,it can adapt to autonomous navigation in mapless environment,and it can be extended to real scenarios for practical application.The extension research provides beneficial exploration for the development of more intelligent bionic navigation technology.