Using Anytime Planning for Centralized Coordination of Multiple Robots in Real-Time Dynamic Environments

We are investigating the use of planning in multi-robot, real-time, dynamic environments. Each individual robot is considered to be an effector of a centralized planning system. In a real-time environment, it is important to be careful to ensure that the time required to find a good plan does not itself reduce the system’s ability to complete its tasks in a timely manner. Our physically situated system coordinates a group of security robots that guard an area against intruders. Our approach uses a variation of the three-level paradigm for robot architectures (e.g. (Bonasso et al. 1997) (Brill et al. 1998)) in conjunction with an anytime approach to planning (Dean et al. 1995) (Zilberstein & Russell 1993). In this domain, the system must ensure that the total amount of time spent planning and then executing the plan is short enough that the security robots can still capture the intruders. Too much time spent planning prevents the security robots from achieving this objective. Our system uses anytime planning to ensure that plans are found in a timely manner. Anytime algorithms return results with a quality proportional to the amount of time the algorithm had available for execution. Our planner initially assigns patrol routes to the security robots that ensure sensor coverage of the area. As intruders enter the area, our planner assigns some of the robots sequences of intruders to pursue, and plans patrols for the remaining robots to maintain sensor coverage. Once tasks are assigned, the security robots carry out their tasks reactively (Brooks 1986) (Brill et al. 1998). Once reactive execution begins, the security robots act independently of the planner until they are assigned new tasks. Each robot reports the current locations of any intruders sensed to the sequencer. The sequencer determines if reassignments are necessary, and if so, reinvokes the planner. Task reassignments are made when new intruders are detected or a security robot is out of range to capture an assigned target. An anytime approach is used by our planner for robot task replanning. For an anytime algorithm to be effective, it is important to specify a progression of results that the algorithm will return. Our progression is based on the concept of plan locality. That is, when a plan fails and needs to be replanned, much of the plan may still be valid and useful as is, and repairing the failure may be sufficient. On the other hand, local failures in a plan can have non-local implica-