Task Understanding For Service Robots

Nowadays,service robots are becoming the focus of research in the field of robotics over the world.Intelligent service robots are regarded as the cutting-edge technology in following 15 years in China.The industry of domestic service robot,whose future market is huge,continues to develop.However,the ability of task understanding and planning in human-robot interaction is still a major bottleneck affecting robots’ services to humans,and there is a huge gap between the local knowledge of robots and the tasks expressed in natural language.(1)the expression of the user instruction is not only a natural language form but also an abstract expression which is under-specified;(2)knowledge gaps between users and local knowledge of robots;(3)map user tasks to a sequence of primitive actions.Therefore,this paper solves the problems of these three aspects,and maps user tasks to a sequence of primitive actions that robot can perform to solve the problem of task understanding.First,the solution to the problem of "the expression of the user instruction is not only a natural language form but also an abstract expression which is under-specified".Because user tasks are expressed in natural language,robots need to translate them to formal representation.However,the user language is rich,it usually uses a different vo-cabulary and will cover most of the daily vocabulary we use.At present,the common practice is by marking a large number of sample data as a training set,and with the deep learning or machine learning method to train a semantic parser.However,marking a large number of task training sets is time-consuming,while related user task datasets are also limited.In order to solve this problem,this paper focuses on training reliable semantic parsers from a limited set of annotations.At the same time,a new heuris-tic method is proposed to recover semantic roles of the current task from the context information of the user task.Secondly,the solution to the problem of "knowledge gaps between users and local knowledge of robots".when a user task sentence is translated into the frame semantic representation,robots need to define the semantics of common verbs to fill the knowl-edge required for task planning.However,most of them depend on hand-coded ac-tion effects.In order to solve this problem,we extracted semantic roles of the formal definition of common verbs from dictionaries or other similar resources,while these definitions can be handled directly by robots.As the existing action description lan-guage did not give the complete description of this type of knowledge,we propose a new meta-language framework as the representation framework of the common verbs and user tasks.Based on the above framework,semantic information of common verbs is extracted from the dictionaries and rewritten in the meta-language framework.Such rewritten knowledge can effectively fill functional knowledge in task understanding.Finally,the solution to the problem of "map user tasks to a sequence of primitive actions".When the semantic information of the user task is obtained from the meta-language,since these metalanguage is an intermediate language representation or can not be directly used by the robot,it needs to be first translated into a logical form lan-guage(ASP rules),together with local knowledge of robots and the semantics of user tasks,input to the ASP solver to generate a sequence of primitive actions.This action sequence is finally executed by the robot in turn.So in this paper,we use the semantic information of FrameNet and other seman-tic dictionaries to fill the functional knowledge of common verbs.At the same time,combined with natural language semantic parsing and answer set programming(ASP)technology to improve the ability of user task understanding.Then,the proposed sys-tem was applied to our " KeJia" robot and participated in RoboCup@Home test.In the experiments,we evaluated our approach using common benchmarks on service tasks and showed that it can successfully handle much more tasks than the state-of-the-art solution.Notably,we deployed the proposed planning system on our service robot for the annual RoboCup@Home competitions and achieved very encouraging results.The main contributions of this paper include:· For solving the problem of "underspecified and limited sets".This paper pro-poses to utilize argument typed dependency features based log-linear model to learn a semantic parser from a limited dataset.Moreover,this paper proposes a nearest neighbour and default principle to recover missing roles from the context of an instruction.Implementing a semantic parser that can handle larger-scale task sentences.· For solving the problem of "knowledge gaps",this paper introduces a meta-language to formalize functional knowledge and procedural knowledge,fills the gaps be-tween local knowledge of robots and user tasks.· For solving the planning problem,this paper combins meta-language and ASP planning,proposes series of algorithms and a new system architecture for task understanding.This increases the degree of autonomy of service robots.