A Review Regarding Deep Learning Technology in Mobile Robots

Abstract

-Deep Learning usage is spread across many fields of application. This paper presents details from a selected variety of works published in recent years to illustrate the versatility of the Deep Learning techniques, their potential in current and future research and industry applications as well as their state-of-the-art status in vision tasks, where their efficiency is experimentally proven to near 100% accuracy. The presented applications range from navigation to localization, object recognition and more advanced interactions such as grasping. Keywords: deep learning, neural network, RNN, CNN, mobile robots I.INTROTUCTION Neural network systems have proven to give results in many types of applications in the last few years, becoming an essential technology in areas involving computer or machine vision. Technological advances which include cheaper hardware, better hardware performance and possibilites, as well as the rise of parallel processing on GPUs allow neural network-based architectures to be employed more commonly in both industry and IT projects, thus to be studied, optimized and refined as a current emerging technology and future base for many products and services.Deep learning refers to the use of neural networks for learning desired features on labeled sample data, in order to be used later for identifying those features in new data. The types of neural networks usually involved have many hidden layers, being called deep, hence the term Deep learning. II.NAVIGATION Mobile robots, as well as industrial robots, have seen succesful use of deep learning in recent years. In [2] the authors explored an architecture with separate neural networks for perception and control. It was used for navigation purposes and tested on a mobile robot with two wheel differential drive. The model was inspired by a DQN (Deep Q-Network), but split into separate networks to create a simplified model.The perception network is a CNN with three convolutional layers. It receives depth information as input and outputs feature representations which are fed into the control network. The latter is built using three fully-connected layers. The authors of the paper motivated the use of a second network for control in order for the robot to rapidly adapt to new environments without pre-training in the new environment. A simulated experimental setup was used involving the use of Gazebo for the 3D environment and robot, as well as a CNNbased reinforcement learning control framework.In order to train the robot, it was made to explore the environment while being remote controlled, thus labelling the depth image acquired from the sensor with the control commands. The trained model was then used to extract feature maps in real-time from sensor data, which were the output of the last ReLU layer. Then the control network was used for estimating the Q-value like in a DQN.The depth image in this case could be regarded as the state s t and it would be memorized along with its action a t , reward r and next state s t+1 which is a new set ...