An Action Language BC-based Architecture For Service Robots

Research on automated general service robot has drawn much attention in recent years. Service robot as a highly intersecting research area, involves work on mechanical engineering, motion control, perception, human-robot interaction, knowledge represen-tation and reasoning, etc. As a result, most research on service robots maily focus on specific abilities rather than the whole robot system. Compared to industrial robots, service robots differs mainly in the following three aspects.First, working environment of industrial robots are mostly fixed, the robots them-selves are designed and programed to fit the specific conditions and tasks. Once run-ning, such robots would be continuously working on the production line 7x24 hours a week with minimum human interference. Meanwhile, any disturb from outside object or human beings would probably cause severe accidents, since industrial robots are usu-ally very powerful but with a minimum detection for outside environment. Upon any accidents, staff with professional skills are called to reset or repair the robots, which cost is non-trival. On the other hand, service robots cannot assume a static nor similar working environment:layouts of our apartments may differ significantly, human be-ings or pets at home bring uncertainties. So service robots must respond to dynamic changes and uncertainties in such a way that they are able to finish the tasks despite all those challenges, or report to human why they fail.Second, perception abilities of industrial robots are mostly limited to their mechan-ical parts rather than environments. Although some newly developed industrial robots also have vision or other sensors to detect and localize interested parts, such sensing is still much limited. In comparison, service robots cohabit with human and it is impos-sible that the robot can ignore any part of the environment:it may just hurt someone. Meanwhile, during the execution of tasks, service robot must be able to identify the changes of the environment and interested objects, in order to adapt to changes rapidly.Third, the interface between human and industrial robots are almost the same as computers. As stated before, the predefined working conditions make the industrial robot minimum need for adjustments during working; such interface are mostly needed before industrial robots start working. But the nature of service robots makes human-robot interaction a basic need, from simple chatting, to obtaining orders from human beings, and seeking for help if anything went wrong.The goal of our research is to develop a general framework for autonomous service robots, with an implementation on realistic environments to furture study the underlying problems. During this work, we recognize two kinds of challenges on developing such a service robot. One kind is from developing the different abilities of the robot, e.g the mechanical parts including base and robot arms, control module which drives the hardware and navigates arms. perception abilities including obstacle avoidance, SLAM and vision. Another kind of challenge is how to integrate all such components to make the robot a running system, adapt to dynamic changes and behave intelligently. Our work focus on the latter one.The proposed framework in this paper is a possible approach to solving the sec-ond challenge. The robot equipped with such framework is able to finish tasks in do-mains that contain incomplete information, underspecified goals and dynamic changes. Human-robot interaction, sensing actions and physical actions are uniformly formal-ized using action langues BC, an answer set programming based action language. After given a task, an answer set solver is called to generate plans that can drive the robot proactively acquire task-oriented knowledge, and execute actions to achieve such task. During the execution, the robot continouously observes the environment and interpret usefull sensing information to use contingent knowledge to adapt to unexpected changes and faults. Our work features three novelties.First, we propose a general modeling methodology of using action language BC, whose semantics is based on answer set programming. BC formalizes transition sys-tems, in which world state and robot’s belief state are represented in a uniform way and how HRI, sensing and physical actions affect them. When a goal is given, even if it is underspecified, plans including proper HRI and sensing actions can be generated to achieve the goal.Second, we extend the traditional "plan-execution-monitor-replan" execution loop for robot plan execution, by introducing a parallel continuous observation mechanism, which catches contingent knowledge. When the robot is carrying out a plan, the sensed results are continuously updated and grounded into symbol representations, and the robot performs reasoning to validate current plan in presence of newly updated infor-mation, so it can adapt to changes rapidly. Such contingent knowledge also helps in consecutive tasks because each time the robot finishes a task, it learns more about the domain, which helps in future tasks.Third, when HRI and sensing actions are treated equally in the execution loop, it brings the challenge of grounding execution results into "epistemic states". For in-stance, when Alice asks the robot to bring her the water from dining table, but after robot reaches the dining table, the perception module finds the water unreachable for the robot arm, or just cannot find any water at all, then how should we represent the sym-bolic representation? Or what is the state when an object detected earlier mysteriously disappeared? By a proper division of different kinds of fluents, we demostrate how to handle symbol grounding for various realistic situations with unexpected changes.The proposed framework in this paper is implemented in KeJia service robot, tested and benchmarked on a modified "General Purpose Service Robot" scenario pro-posed by RoboCup@Home competition. KeJia service robot has been participating in RoboCup@Home since 2009, and has won a champion in 2014, Brazil, several runner-ups in 2011,2013 and 2015.