A spatial cognition approach based on grid cell group representation for embodied intelligence
Navigation,both in biological organisms and in artificial systems,is a complex process that requires an understanding of the environment and the ability to plan a route through it.While traditional navigation algorithms depend heavily on structured scene modeling,mammals demonstrate a more robust approach.They generate cognitive maps using a combination of different neuron types in the entorhinal cortex and hippocampus and thus achieve effective navigation through unfamiliar or unstructured terrains with ease and robustness.Therefore,the robustness of mammals in navigating through diverse environments suggests that a brain-inspired approach may offer superior performance in unstructured or dynamic terrains,particularly for mobile robots.Recent progress in cognitive neuroscience and artificial intelligence has shed light on the potential benefits of brain-inspired navigation.Given the benefits,there has been an increasing interest in applying insights from neuroscience to robotic systems.This fusion of fields is not only expected to enhance the capabilities of robots but also provide a deeper understanding of the mammal brain's spatial cognition mechanisms.In the presented study,we propose a novel cognitive map construction system tailored for mobile robots based on a grid cell group representation model,combining spatial cognitive mechanisms of the brain with simultaneous localization and mapping(SLAM)algorithms.In this system,the grid cell and place cell model collaborate to accomplish place representation and path integration.While the place cells are crucial for representing the spatial layout of the environment,the grid cells function as a spatial metric and perform path integration.In the grid cell group representation model,the grid cell group is characterized by a high-dimensional vector and the movements in two-dimensional space are represented using a matrix.Compared to the continuous attractor neural network(CANN)model,the oscillation interference(OI)model,and models based on recurrent neural networks,this model naturally generates hexagonal grid-like firing patterns and offers a clearer geometric and algebraic interpretation.The combination of grid cells and place cells model allows the robot to understand its surroundings and its location more effectively.To validate our proposed system,it is verified using publicly available real-world datasets and the results were promising.The metric characteristics of the grid cell group model were clearly observable,confirming its suitability in representing space in a manner analogous to mammalian brains.More importantly,the experiments showcased the system's capability in constructing a cognitive map.Beyond its application in robotic navigation,our brain-inspired design offers another advantage.It presents an opportunity to delve deeper into the spatial cognitive mechanisms of the mammalian brain.By creating models that emulate these mechanisms,we not only propose a novel cognitive map construction system but also pave the way for a more comprehensive understanding of mammalian cognitive processes.In conclusion,our study underscores the potential of a spatial cognition approach based on grid cell group representation.By combining insights from neuroscience with modern robotic algorithms,we are a step closer to achieving truly embodied intelligence of mobile robots.
cognitive map constructionbrain-inspired navigationembodied intelligencegrid cellmobile robotsimultaneous localization and mapping
NETL
NSTL
万方数据
维普
