Research On Key Technologies Of Indoor Robot Navigation Based On Deep Learning

Autonomous navigation is the basic function for indoor mobile robots,which is also the premise and key technology to build fully autonomous and intelligent robots.With the development of robot-related technologies and the expansion of robot applications,robots are designed to deal with more challenging tasks and their working environments are increasingly complex and diversified,thus bringing new requirements and challenges for the navigation technology.The indoor robot navigation systems not only need to maintain the accuracy and stability in fixed scenarios,but also need to be able to deal with the dynamic changes in the environment.In the co-existing environment with humans,safety and the impact on human comfort should also be considered.This project takes the indoor mobile robot platform as the research object and makes use of the existing research results in the field of deep learning to construct robot navigation systems based on different learning paradigms,with the aim to introduce the neural network based machine learning algorithms into the research field of robot navigation.Whiling solving robot navigation tasks through learning methods,three different kinds of learning paradigms are used,namely supervised learning,reinforcement learning and unsupervised learning.These algorithms are used individually or in a combination way to solve practical problems and to improve learning efficiency.Compared with the traditional SLAM-based navigation system,the learning-based navigation system is more flexible and expansible,emphasizing to endow the robot with intelligent perception ability and navigation strategy learning ability.This enables the robot to learn and explain the world as humans,so that the robot is able to deal with navigation problems in complex environments.Compared to the traditional navigation systems,which mainly focus on metric information but lacks semantic information,the proposed vision perception system is able to maintain both types of information.It is based on the existing research achievements in the field of computer vision and greatly improves the robot's perceptual intelligence.For the indoor scene classification,a deep neural network based solution is proposed,where the local image scene and global geometric information are combined together to improve the classification accuracy.Aiming at the improving the detection accuracy of small-scale objects in object detection tasks,an improved scheme is proposed for the existing detection model and the attention mechanism is used to improve the position accuracy of the detection result for small objects.For this aim,a distance-based attention mechanism is proposed and is able to improve the detection results of small objects without adding any computing overload.Finally,a goal-reaching system based on the object detection is proposed,which enables the robot to move to the target area according to its visual information.For navigating in dynamically changing environments,an hierarchical strategy that combines imitation learning and reinforcement learning is proposed,which enables the robot to move to the target position in environments with changing factors.The standard imitation is modified to support goal-directed movement,where the current observation and goal representation are combined to learn the action policy.A global planner based on the modified imitation learning is used to implement global planning in static environments.To deal with the spatial changes in the environment,a local planner based on modified reinforcement learning is built to implement local obstacle avoidance.The local planner is trained in simulation environments and then transferred into to real world.Through switching between these two navigation strategy,the robot is able to complete goal-directed navigation in dynamic environments.Considering that the SFA algorithm is able to extract spatial-related features from image sequence in an unsupervised way,a visual navigation system is proposed,which is able to solve navigation tasks in a self-organizing fashion.It is based on the SFA and self-organizing networks,and involves topological map building,self-localization,orientation detection.Through combining with an action memory module and an action output module,which are built based on different networks,the proposed system is able to perform navigation based purely on the vision system,without the need of any prior knowledge and human designs.To aromatically generate the hierarchy in hierarchical reinforcement learning,a solution which combines the learning result of SFA and self-organizing networks is proposed.The SFA algorithm is used to learn spatial features from visual images and the self-organizing network is responsible for building a topological map of the environment in an unsupervised way.This produces a nature hierarchy of the environment.Based on this,two action policies that work on different levels of abstraction are learned,where the high-level one is used to perform route planning and the low-level one is used to perform goal-directed navigation.In this way,the robot is able to learn navigation strategy in large-scale environment with high efficiency.