Integrated Task and Motion Planning for Mobile Service Robots

I. INTRODUCTION " Planning " , or selecting a sequence of actions to achieve a goal, has been a core focus of interest within the field of Artificial Intelligence (AI) since the field was founded in the 1950's. Initially, the focus of attention was on task planning which is concerned with sequencing actions within a symbolic representation of the state space [2]. When action costs are further incorporated into the planning process, existing task planners can find the optimal plan that minimizes the overall plan cost. A largely independent thread of research has arisen on motion planning that focuses on producing a continuous motion while avoiding collision with obstacles in 2D or 3D continuous space [8]. Traditionally, motion planning has been concerned with computing a path connecting a single start configuration to a single goal configuration, without any concern for sequencing of subgoals. Task and motion planning remained generally independent in large part for a long time, because until recently, physical robots have only been able to execute very short missions that could be solved entirely with motion planning algorithms. Meanwhile, complex task plans have needed to be generated in simulation or in purely software domains. However, with the recent advances in long-term autonomy on mobile robots in large-scale indoor environments [10, 6], there is a pressing need for the ability to generate task plans that are fully aware of, and indeed dependent upon, the grounded costs of actions that can only be determined by motion planning algorithms. In principle, motion planning costs could be evaluated on all the possible action sequences in the task planning space. However , especially in cases with combinatorially many possible sequences (for example if there are many possible separate places to buy coolers, ice, milk, and hot dogs), doing so is computationally infeasible. The aim of this research is to fully integrate task and motion planning in order to find the lowest cost, optimal plan in task planning 1 in a computationally tractable manner. We develop a novel task and motion planning algorithm that has bi-directional communication between task and motion planners: the task planner is capable of computing a symbolic plan with the lowest cost conditioned on existing evaluations of action costs; and the motion planner is capable of evaluating the true cost of these constituent actions in the " lowest-cost " plan. This interactive process is repeated until a lowest-cost