Research On Task Execution-Oriented Environment Modeling Problem For Home Service Robots

Driven by social demand growth and technological development,mobile robots are gradually entering the families to provide users with household services.The environmental modeling,as the basis for the robot to perform household services,integrates the core technologies of environmental perception,understanding,and representation.However,when the robot operates in an open,dynamic,and unstructured home environment,it still faces many challenges,such as navigation safety,task execution efficiency,and long-term autonomy.In this context,how to build the accurate model of the complex home environment to support the robot to perform household tasks safely,efficiently and autonomously for a long time is the key to promoting the entry of mobile robots into the families and the realization of robot intelligence.In this dissertation,guided by the needs of mobile robots to perform household tasks in the home environment,the task execution-oriented environment modeling problem for home service robot is investigated deeply,and the research products are as follows:1.In order to enable the robot to perform household tasks safely and effectively,the problem of collision between the robot and the spatial obstacles in the environment is studied,and a two-dimensional grid map construction method towards robot obstacle avoidance in three-dimensional space is presented.First,a visual sensor is introduced to detect spatial obstacles that are not fully visible to the laser sensor,such as tables and chairs,and then we propose an approach to convert the visual information into two-dimensional pseudo-laser data representation.This pseudo-laser data representation effectively describes the spatial obstacles in the environment.Besides,based on the laser data and pseudo-laser data,a grid map fusion algorithm is developed to generate an improved two-dimensional grid map that can fully represent the obstacle information,which solves the problem of incorrect representation of obstacles on the traditional grid map.On this basis,by weighing the laser and pseudo-laser data,a three-dimensional obstacle avoidance strategy for the robot is presented to avoid collisions between the robot and spatial obstacles in the environment.Experimental results show that the improved grid map together with the presented obstacle avoidance strategy allows the robot not only to plan a‘real’collision-free path,but also to navigate safely and reliably in both static and dynamic scenarios,and the safety and robustness of robot operation are significantly improved.2.To enable the robot to complete household tasks efficiently,the problem of fast locating of target objects guided by a prior knowledge is studied,and a metric-topological map construction method towards object search is proposed.First,based on the reference of objects,we build a novel metric-topological map,in which the topological nodes refer to the object-related locations in the environment.The built map not only removes redundant nodes that are irrelevant to objects,but also facilitates the robot to locate the target object.In order to effectively create topological nodes on the map,a rapid adjustment approach of the robot viewing angle is designed.In addition,inspired by the behavior of humans searching for objects,by considering the robot’s current position and the distance from the robot to the location where the object is most likely to be found,a priori knowledge-based object search strategy is proposed,in order to allow the robot to allow the robot to preferentially search for the most promising location,thus increasing the search efficiency.Experimental results demonstrate that the proposed solution can enable the robot to find the target object efficiently and reliably.3.For efficiently supporting the robot to perform object manipulation,the problems of low calculation efficiency and large memory requirements in three-dimensional environment representation are studied,and a local three-dimensional representation method towards robot manipulation is presented.First,based on the laser data,a two-dimensional grid map is built for robot navigation.Based on the pruned visual observation data,a task-oriented three-dimensional representation approach of the environment is developed for collision detection during object manipulation.Then,the environment modeling towards robot manipulation is realized by innovatively integrating the two-dimensional grid map and local real-time three-dimensional environment representation,which enables the robot to construct a local three-dimensional environmental model in real time only when it needs to manipulate objects,thereby reducing the three-dimensional representation of redundant scenes.Subsequently,a task-oriented object pose estimation method is proposed to guide the construction of local three-dimensional models and realize task-oriented object manipulation.Finally,the presented scheme is fully evaluated in terms of memory requirement,computing time,collision detection,and object pose estimation,and its effectiveness and efficiency are verified through practical application scenarios.4.In order to enable the robot to effectively perform housework tasks in an open and dynamic home environment for a long time,the problem of the robot’s long-term adaptability in a dynamic and uncertain environment is investigated,and a multi-level and multi-granularity semantic environment modeling method towards long-term autonomy of the robot is proposed.First,based on the spatial location relationship between typical objects and rooms,a universal probabilistic model and semantic model are established to improve the adaptability of the robot to the home environment.Then,an interactive association method is designed to maintain the mapping relationship between the probabilistic model and the actual scene.On this basis,based on Bayes’ theorem,a task-driven reasoning and updating mechanism is developed to guide the robot to preferentially search for the space where the target object is most likely to be placed.Besides,a hierarchical and different granularity representation approach is used to realize the construction of multi-level and multi-granularity semantic environment model.In addition,the created environment model can be shared with other robots to perform housework tasks.Extensive experiment results show that the presented solution allows the robot to find the target object autonomously and efficiently for a long time,while realizing human-like search performance.