Research On Key Technologies Of Autonomous Human-Robot Interaction For Education And Training Of Children With Autism

In recent years,with the continuous development of robot technology,its application fields have been expanding.Using intelligent robots to reduce the symptoms of children with autism through human-robot interaction is a hot issue in current research.Studies shown that robots can increase the engagement of children with autism in tasks of education and training,and stimulate their social behavior newly.However,in the face of tasks of education and training for children with autism,autonomous human-robot interaction is a very challenging problem.This paper takes autonomous human-robot interaction as the main line,and aims to serve the education and training of children with autism.The research focuses on the control architecture design of robots,the engagement evaluation of interactive objects,and the decision-making of robots.The main work as follows:1)The control architecture of the robot is studied for the education and training of children with autism.The control architecture regulates the relationship between the modules of the robot and the direction of information circulation.It is the basis for the robot to achieve autonomous human-robot interaction.Firstly,the traditional methods of education and training for children with autism are studied,and the two behavioral intervention methods of "Discrete Trail Teaching" and"Floortime" are analyzed.The model of human-robot interaction suitable for the robot's intervention is proposed.Then,the control paradigm of the robot is analyzed,and the characteristics of the robot for the autonomous human-robot interaction are given.Finally,based on the cognitive model of ACT-R,the traditional strategies of education and training are integrated into the robot's paradigm of perception-cognition-action,and the control architecture for education and training of children with autism is proposed.2)The engagement evaluation of interactive objects in the process of autonomous human-robot interaction is studied.Starting from the role of social signals in the task of engagement evaluation,the process of the engagement evaluation for education and training of children with autism is analyzed.On this basis,based on dynamic Bayesian Network,the topological structure of engagement evaluation model is constructed.Considering the fact that there are few training samples,the prior knowledge of domain experts is fully utilized to parameterize the model.In this process,the qualitative descriptions of domain experts for related problems are transformed into the numerical solution of the engagement evaluation model.The main steps can be described as follows:firstly,the linguistic variable set is established and the qualitative opinions from experts are collected;secondly,linguistic variables are fuzzified and experts' opinions are fused;finally,exact values are obtained by using the method of integral values.The validity of the engagement evaluation model is verified by simulation examples.3)Considering the individualization of interactive objects in the process of human-robot interaction,the engagement evaluation is studied to improve its generalization ability.Firstly,the influence of individualized behavior of the children with autism on engagement evaluation is analyzed,and the uncertainty in knowledge is expressed based on rules.Then,the uncertainty of knowledge is expressed by the belief rule base,and the conversion relationship between confidence and probability distribution is given.On the basis of the study of engagement evaluation in dynamic environments,a model of engagement evaluation based on the hierarchical belief rule base is established.Finally,the objective function and constraints are set,and the relevant parameters in the engagement evaluation model are optimized by differential evolution algorithm.The adaptability of the proposed method to the personalization of interactive objects is verified by a simulation example.4)Based on the characteristics of autonomous human-robot interaction in the tasks of education and training for children with autism,the decision-making of the robot is studied.Firstly,aiming at the tasks of education and training for children with autism,the decision-making framework of the robot is established based on reinforcement learning.Then,the decision process of the robot is described mathematically,and the goal of the decision-making is defined.Finally,a hierarchical decision-making method for the robot is proposed.Among them,in the high level,the target task in the current environment is formulated based on imitative learning,and in the low level,the action selection of the robot is realized based on interactive Q-learning.Based on this,the external feedback information is fully utilized,and the processing mechanism for feedback information is studied.In this way,while improving the learning efficiency of the robot,the negative influence of the error of the feedback information on the robot's decision-making is suppressed.5)The experimental study of practical application is conducted for tasks of education and training for children with autism.Based on the humanoid robot NAO and the design of a typical scheme of training,the engagement evaluation method of interactive objects and the strategy of the robot's decision-making proposed in this paper are verified by practical application.