Research On The Obstacle Avoidance Of Mobile Robots Based On Deep Learning

In recent years,with the development of artificial intelligence,especially deep learning,it has become a research hotspot to realize robot obstacle avoidance through self-learning.To realize autonomous learning is an important step to realize intelligent robot,Autonomous learning,an important step to realize intelligent robot,is beneficial to improve the robot's behavior strategy,adaptability,and robustness in unknown and complex environment.The convolution neural network,a kind of supervised feature learning method,can learn the corresponding characteristics from large-scale data.By combining the deep convolutional neural network,the mobile robot is endowed with the mobile brain to realize intelligent perception and decision-making.In view of this,this paper studies the application of deep learning in the field of control engineering,and performs a systematic study on the obstacle avoidance of mobile robots by deep learning.The main research work of this paper was as follows:(1)The mobile robot platform was built based on the robot operating system(ROS)to drive chassis,realize remote control and image display.(2)Robot obstacle avoidance algorithm was designed based on end-to-end learning.This study proposed an improved model based on the AlexNet foundation network,and the network training platform was built by combining with the Caffe,which is a deep learning framework.The network model was trained to use images as input,and the CNN network outputted predictive commands,including turn left,turn right and go straight.We analyzed the results by combining training curve,characteristics of the visualization and sample test.Furthermore,the generalization ability of the model was tested to verify the validity.(3)The trained neural network model was transplanted into the mobile robot platform by ROS_Caffe.By combining the deep learning model with the robot operating system(ROS)through ROS_Caffe,it is convenient and quick to realize the connection between the soft hardware of the robot and deep learning.Additionally,the Web interface was set up to view the input image and model prediction output of the network in real time at remote terminal through browsers.After tests in actual environment,the observed environment information of the robot was processed by the trained model to improve the ability of outputting predictive commands.The results showed that the robot could plan reasonable routes to avoid obstacles,with a high success rate in the experiment.And the test results of the other obstacles showed that our training model had certain generalization ability.