Cooperative Team Plan: Planning, Execution, and Replanning

In the context of robot team control, this paper focuses on a framework for designing a team plan and its projection onto individual robotic agents. A mission plan, represented as a coloured Petri net, is calculated through constraint optimisation from goal and time requirements. The mission plan is then turned into a hierarchical team plan through reduction rules that also structure the dynamic hierarchical team organisation. Hence each level in the team plan corresponds to an abstract plan at the corresponding subteam level in the team hierarchy. Controlling an agent individually requires extracting individual information, such as activities involving the agent as well as interacting agents or subteams at each subteam level: the team plan is projected onto individual agents. At runtime events may disrupt plan execution. A reaction is executed while diagnosis triggers replanning which is performed as locally as possible. Introduction In the agent world activity planning has been widely studied. The increasing complexity of the jobs assigned to agents, especially robotic agents, has led to using groups of agents. The groups, when organised and aware of their organisation, are called teams. The problem of team planning is considered difficult (state-space size of (2−1)k! ∏k j=1 uj , with m the number of agents, k the number of goals, uj the number of recipes for the jth goal). The general framework is a mission specified in terms of objectives: physical robotic agents are operated in order to carry out the objectives and they are hierarchically organised in a team. As outlined later on, most architectures currently proposed either do not take advantage of the agents being designed to operate as a team or restrict the agents to use reactive behaviours. Replanning itself is often considered as a separate problem. This paper aims at emphasizing the relationship between the team plan and individual agents’ plans and formalising and integrating the replanning process through the use of Petri nets (PN — see Appendix) (BonnetTorrès & Tessier 2005a; 2005b). In the first section the mission plan and team organisation are derived from initial problem data. Then the plan is broken down into individual executable plans. The next section Copyright c © 2006, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. details the execution of the plan and the replanning process. Finally related works are presented. From Mission Requirements to Mission Plan Mission Plan Construction The objective of the mission is decomposed into a hierarchy of goals to be carried out. The leaves in the hierarchy are elementary goals. Executing a task corresponds to satisfying an elementary goal. Recipes (Grosz & Kraus 1996) give courses of actions for performing the tasks. Several recipes requiring different sets of services may be available for the same task allowing to achieve it – and performing the associated goal — in a number of fashions.