| With the booming development of robotics and artificial intelligence technology,mobile robots have made gratifying achievements in many industrial scenarios.More and more common intelligent robots,such as storage robots,security robots,medical robots,and service robots,are gradually integrated into our daily lives.In the rich daily life applications of mobile robots,the research on mobile robots in indoor scenes has always been a hot topic in this field.The navigation decision algorithm is the upstream basis for mobile robots to complete complex tasks in indoor scenarios,and its navigation accuracy and reliability greatly impact the function and application of mobile robots.Compared with the fixed task and monotonous working environment of industrial robots,the indoor living environment puts forward higher requirements on the navigation decision-making ability of mobile robots.Therefore,it is of great scientific significance and social value to study the navigation problem of mobile robots in the indoor environment.This thesis focuses on the navigation decision-making problem of mobile robots in indoor scenes,takes the precision degree of navigation instructions obtained by mobile robots as a clue,and studies mobile robot navigation decision-making problems from the perspectives of precise instructions and natural instructions.In the mobile robot indoor navigation task based on precise instructions and environmental priors,it is challenging to balance the system cost,real-time performance,localization accuracy,no cumulative error,and system scalability,while the existing system mainly focuses on obtaining precise position information.In mobile robot indoor navigation tasks based on precise instructions and unknown environments,the implementation difficulty of end-to-end collaborative navigation tasks in a mapless environment lies in the reliability of the decision-making model and its adaptability to collaborative scenarios.In indoor navigation tasks based on abstract natural instructions,it is a great challenge to effectively understand abstract language instructions and build a navigation decision model.In response to the above challenges,this thesis introduces a localization and navigation system based on precise location information,a mapless collaborative navigation model based on laser perception and reinforcement learning,and a vision-and-language navigation model based on natural instructions and multi-modal learning.The research contents and main contributions of this thesis are as follows:1.In view of the problem that the existing localization and navigation systems cannot take into account the performance of system cost,real-time performance,localization accuracy,no cumulative error,and system scalability when obtaining precise position information,a mobile robot localization and navigation system based on a modular sonar array is designed and constructed independently in this thesis,and we further verify the performance of the system.This system can provide centimeter-level precise position information for mobile robot localization and navigation tasks based on precise instructions and has several advantages such as high localization accuracy,low cost,no cumulative error,and convenient deployment.At the same time,an initialization calibration method based on an external SLAM system is proposed for the scalability and deployment convenience of the system.2.For the indoor mapless navigation task of mobile robots based on precise instructions,a deep reinforcement learning framework is introduced to construct the navigation decision-making agent of mobile robots.It can solve the single mobile robot end-to-end navigation tasks and the multi-robot cooperative cases.This thesis proposes the Parallel Deep Deterministic Policy Gradient(PDDPG)algorithm by effectively utilizing the mobile robot laser sensing input and precise position information instruction.The proposed model’s convergence and effectiveness in cooperative navigation tasks are verified in experiments.In this thesis,the navigation decision-making model is trained by constructing the return function and learning course so that mobile robots can avoid various obstacles in the scene,efficiently construct the specific formation required by cooperative navigation,and maintain the specific formation and running rate during the real-time decision-making process of cooperative navigation.3.This thesis proposes a compelling visual language navigation architecture in indoor scenes when we focus on the navigation decision-making research based on abstract natural instructions.By constructing a recurrent architecture to enhance the coupling degree of perception front-end and decision-making back-end,our method combines the multi-modal pre-training model and deep reinforcement learning algorithm framework.Specifically,the proposed architecture adopts the visual language multi-modal pre-training model as the front-end perception model of the agent for multi-modal information perception and fusion.It uses the pre-trained model to build the policy backbone in the back-end navigation decision-making while improving the utilization of trajectory data by building a locus experience pool.In addition,this thesis introduces a proximal policy optimization model to ensure the navigation policy’s stable update and the decision model’s training efficiency.For the above research content,this thesis achieves excellent research results in mobile robot navigation decision-making in indoor scenarios,which effectively complements and improves the exploratory research on mobile robot navigation decision-making. |
