Task Planning And Explanation For Service Robots In Open Worlds

When the robot owns a series of basic capabilities required by a range of service tasks,task planning can integrate these basic capabilities to form an intelligent robot system that can complete a variety of tasks.An important goal of artificial intelligence is to enable an agent to solve planning problems with incomplete information in uncertain environments,and to be able to flexibly respond to environmental changes during the execution of the plan.As service robots enter open worlds,uncertain and incomplete information in open worlds places higher requirements for the capability of task planning.Currently,the capability of task planning is still a bottleneck affecting the capability of service robots in three aspects:(1)uncertain and incomplete information in open worlds undermines the preconditions of classical planning methods,and results in unsuccessful planning;(2)dynamic changes in open worlds may lead to the invalidation of current plan or the emergence of new solutions with fewer costs;(3)classical planning methods fail to explain what prevents the planner from getting a plan.In this dissertation,we address these three issues to improve the capability of task planning,so that the robot can better adapt to open worlds and have better user-friendliness.For planning tasks,there are situations where the planner is unable to obtain any plan either because the robot lacks the required capabilities or because of environmental factors.Classical planning methods are often unable to deal with such situations and obtain corresponding explanations.For such situations,this dissertation proposes to use virtual actions to explain planning failures.A virtual action is a special action with fewer preconditions and only one effect.Like edges in a graph,virtual actions can connect vertices that are not connected.Since a virtual action contains only one effect,when it appears in the plan,we can know which state causes the planning failure.Based on virtual actions,we propose a hierarchical planning framework to avoid the impact of virtual actions on normal planning.An open world is often dynamic,and sometimes robots cannot perceive changes in the open world in real time due to the fact that the robot’s perception cannot cover the entire environment at the same time.The difference between the world perceived by the robot and the real world poses two problems:(1)the lack of necessary information,especially the information that isn’t covered in the instructions but is needed to complete the task,makes it impossible to generate a plan;(2)changes in the world may lead to the invalidation of the current action or the emergence of better solutions.At present,the common practice is to use the "assumption-grounding" method to deal with the first problem,i.e.,to "assume" that there are objectives that meet the task description,and "ground" whether the objectives exist during the execution.The second problem is addressed by using "replanning",i.e.,replanning when the current action cannot be performed or when the world changes.However,there are some problems with these approaches.When dealing with the first kind of problem,the known information of the robot is not considered and redundant "assumptions" may result in missing better plans.Since it is not known which intermediate objects are planned to be employed,corresponding assumptions cannot be made when information about intermediate objects is lacking.When dealing with the second problem,there is a lag in replanning until the current action cannot be performed,and replanning for all perceived environmental changes is over-planning.We propose two approaches to address these two challenges separately:(1)making assumptions by combining the goal and known information and using virtual actions to find the missing intermediate information;(2)establishing equivalent instance rules based on the task description and deciding whether to replan based on the changes of equivalent instances,i.e.,plan strongly related changes.Two objects are equivalent,meaning that they are equivalent in meeting the user’s needs.In summary,this dissertation proposes task planning based on virtual actions and using the virtual actions in the plan to construct the explanation for the initial planning failure,using an extended "assumption-grounding" approach to generate initial states and goals to cope with insufficient information,and using a balanced replanning strategy to adapt to environmental changes.Experiments show that the techniques proposed in this dissertation,such as virtual action,largely improve the robot’s planning and interpretation capabilities in open worlds.The main contributions and innovations of this dissertation include:(1)To address the problem that the existing task planning methods are difficult to explain the planning failure,a new solution is proposed:a new planning method is implemented based on virtual actions.According to the virtual actions in the generated plan,the reason of the initial planning failure is located.The method enhances the interpretability of task planning.(2)To address the problem that the classical task planning algorithm cannot work successfully with incomplete information,an initial state and goal generation method based on equivalent instances,virtual actions and extended "assumption-grounding" is proposed,which enables the robot to successfully carry out task planning with insufficient information.(3)To address the problem that the existing replanning methods can not balance finding a better plan and carrying out less replanning,a replanning method based on plan strongly related changes detection is proposed,which only carries out replanning for plan strongly related changes,and better plans can be detected with less replanning.