Chinese Instruction Parsing Of Home Service Robot Based On Deep Learning

The experience of human-computer interaction is an important standard for measuring the intelligence of service robots.Making robots understand and execute instructions issued by users is a hot direction in the research of service robots.In the process of robot understanding instructions,how to parse the instructions into structured action sequences and map the action sequences into machine instructions that the robot can execute are two important research contents in the instruction parsing work.With the continuous development of deep learning,researchers integrate deep learning technology into the research of instruction parsing and achieve excellent results.However,the performance of existing methods in processing complex instructions with multiple sets of action sequences at the same time is not ideal,which affects the accuracy of instruction parsing to a certain extent.Therefore,this dissertation studies the related issues of generating action sequences and mapping sequences to robot actions from two aspects: improving the accuracy of instruction parsing and simulating.First,for the action sequence generation task of Chinese instructions,proposes a sequence annotation method and an action sequence generation model suitable for multiple sets of action instructions.The EA-BIO labeling method is proposed by improving the existing BIO labeling method,and the ASE-ERNIE model is proposed based on the ERNIE method,which can generate all the action sequences contained in the Chinese instructions under the premise of ensuring accuracy.In addition,since the pre-training model has been pre-trained on a large number of Chinese corpora,it makes up for the lack of Chinese instruction data in the home environment.Secondly,in order to enable the robot to accurately execute the instructions issued by human users,this dissertation maps the action sequences generated by Chinese instructions into machine instructions that the robot can execute.Through the pre-designed machine instruction library,the basic actions that the robot can perform smoothly are collected,and then the action sequence is mapped into machine instructions using a deep learning model based on bidirectional GRU method.At the same time,in order to improve the generalization ability and anti-interference ability of the model,the training of the model is completed by means of confrontational training,and the performance of the model is improved by adding noise to the input vector of the model.Finally,in order to further verify the effectiveness of the method proposed in this dissertation,this dissertation simulates the robot to execute the command of gripping and puting objects.During the demonstration,it mainly involves the design and construction of the simulation environment,object recognition and action planning.Gazebo was selected to build the simulation environment used in the experiment based on ROS platform,and PR2 robot and Move It planner were selected for unified planning of machine instructions,observe the robot’s understanding and execution of instructions.