Research And Application Of Brain-inspired Models For Robot Cognition

Cognitive robots are intelligent robotic systems that empower robots to process information and perform tasks in a human-like perception,learning,and reasoning manner.Enhancing the cognitive capabilities of robots is beneficial for robots to exhibit a high degree of adaptive and self-learning capabilities in dynamic open scenarios to accomplish tasks such as perception,learning,and decision-making,accelerating their applications in different fields.One of the important development directions to enhance the cognitive ability of robots is to build a brain-inspired robot cognitive model based on the information processing mode,connection structure and pulse neuron dynamics properties of the brain.At present,the research on brain-inspired cognitive models is still in its initial stage,and it is still a great challenge to use the computer system inside the brain to enhance the cognitive ability of intelligent robots.At the same time,robots often need to have various basic abilities such as cognition of the environment,associative memory among related things,and decision making operation of objects to accomplish a complex task.Therefore,based on the information processing mechanism of the brain,the connection structure between brain regions and the dynamics of neurons in the brain,this paper investigates the information representation,spatial cognition,associative memory and decision-making operations faced by robots in the process of cognition.The main research contents and innovations of this paper are as follows:(1)A brain-inspired mixed dimensional sensory representation model is proposed.The model incorporates the compressive downscaling and sparse dimensional expansion properties present in the sensory information processing pathways of the brain,and experimental results show that the accuracy achieved using the proposed model is superior to that achieved using either the compressive downscaling or sparse dimensional expansion information representation methods alone.Earlier studies have demonstrated the computational advantages of each of the compressive and dimensional expansion representation methods,but the mixed gain has rarely been discussed.Inspired by the information processing procedure of the biological brain,this paper investigates the effects of compressive downscaling and sparse dimensional expansion on information representation and proposes a mixed-dimensional based sensory representation model.The proposed model consists of a stimulus layer that collects input,an encoding layer that compresses and downscales the stimulus information,a cortical layer that sparsely dimensionalizes the encoding layer,and an output layer that performs a specific task.The coding layer reduces the redundancy and noise of the input information through compression and dimensionality reduction,and retains the more valuable information for transmission to the cortical layer;the cortical layer increases the separability of information through sparse dimensionality expansion and enhances the fault tolerance for intra-class pattern recognition.In this paper,the effectiveness and robustness of the proposed model are verified by quantitative analysis of information representation in different layers of the model and experimental results in image retrieval and face recognition.In addition,the model can be adapted to the algorithm of each layer in the model according to the needs of the task,so it can be used as a general information representation framework and helps to revisit how the neural pathways integrate input signals,process information,and make decisions based on them.(2)A robot spatial cognition model based on the local sensitivity of the fruit fly olfactory pathway is proposed,which can effectively improve the efficiency of loop closure detection during the spatial cognition of the robot.The the spatial cognition inspired by entorhinal cortex-hippocampus circuits requires loop closure detection for robot relocalization and accumulated error correction.Current brain-inspired robotic spatial cognition systems often use hand-craft features and brute-force search strategies for loop closure detection,which are not well suited to solve the problem in challenging(dynamic scenes,lighting changes)and large-scale environments.Based on the idea of the mixed dimensional sensory representation model,a loop closure detection method with mixed depth features and brain-inspired local sensitive hash(LSH)is proposed.Among them,the depth features are obtained based on the learning of a large amount of image data,which can effectively reduce the influence of illumination and pose on the loop closure detection and improve the accuracy of similar scene detection results.Meanwhile,the method uses multiple hash tables to speed up the loop closure detection.(3)A bidirectional associative memory(BAM)model based on spiking neural network(SNN)is proposed to establish relationships between actions,objects and their attributes,and the model is deployed on a humanoid robot to infer user intent and improve the human-robot interaction experience.Neuroscience research has identified several memory-related regions in the brain(including hippocampus and entorhinal cortex),and memory is the brain’s summary and storage of experienced events,which is the key component to achieving intelligence in animals.Therefore,establishing associations and saving memories between objects in a cognitive environment can help intelligent robots better understand human intentions and improve the human-robot interaction experience.The SNN-based BAM network uses a supervised learning algorithm to complete the association learning between different things.To effectively complete the learning process of the spiking BAM network,this paper also proposes an exclusive or(XOR)phase encoding method based on the index of encoding neurons and the activation state of the input data.This encoding method is also used in experiments to accomplish the recognition of gesture actions.The experimental results show that the proposed BAM network has good convergence performance and accuracy.The human-robot interaction experiments also show that associative memory in cognitive development can help improve the robot’s ability to infer human intentions during human-robot interaction.(4)A SNN-based robot cognitive decision model is proposed for robots to execute their grasping decisions in the workspace.The robotic arm is an important tool for the robot to interact with the environment,and it is an important carrier to reflect its intelligence,among which the grasping operation is the basic operation skill of the robotic arm,and how to use the strong computational power and low-power characteristics of biological neurons to complete the grasping operation is one of the research directions of robot cognitive decision making.At present,SNNs are difficult to train in complex tasks,and the artificial neural network-spiking neural network(ANN-SNN)conversion algorithm solves the problem by the frequency equivalence principle,so this paper implements the grasping operation of a robotic arm based on the ANN-SNN conversion algorithm.In this paper,an ANN model for grasping pose prediction is first proposed and the training of this network is completed on the grasping dataset.The trained ANN is then transformed to obtain a deep SNN so that the robotic arm can rely on the SNN for cognitive decision making to complete the planar grasping operation.During the network conversion process,the spiking frequency of neurons in the SNN is increased by the channel-wise normalization method,which reduces the accuracy loss after conversion.