Integrating skills into multi-agent systems HOLGER FRIEDRICH, OLIVER ROGALLA and RU È DIGER DILLMANN Institute for Real-Time Computer Systems and Robotics (IPR), University of Karlsruhe, D-76128 Karlsruhe, Germany Received February and accepted October 1997 Currently, an important topic of robotic research is the design and development of multi-agent robot systems (MASs). In these a number of autonomous robots cooperate and coordinate themselves in order to pursue given goals. The agents of an MAS not only have to work autonomously or in cooperation with other agents, but in dynamic, relatively unstructured environments. Therefore, the agents require agent-speci®c but ¯exible skills to cope with their tasks and the environment's variability. On the other hand, the actions to be performed by agents in an MAS have to meet certain requirements imposed by the MAS's structure. The representation of actions has to support planning, inter-agent communication, task negotiation etc. In this paper, we describe a method of combining the agent-speci®c nature of skills with the requirements for a general action knowledge representation inherent to MASs, by presenting elementary operations (EOs) that provide an appropriate interface. Ó 1998 Chapman & Hall Keywords: Robotics, multi-agent systems, skills, knowledge representation 1. Introduction Today, the need for ¯exible manufacturing systems in- creases due to economic and ecological constraints. These demand a shorter time-to-market, production of small product series with high component variability, and further automization of tasks that are currently performed by human workers. With respect to these requirements, cen- tralized, monolithic control architectures of industrial ro- bots and manufacturing devices have certain shortcomings because of their inherent in¯exibility. Throughout the last few years the design of distributed systems of autonomous agents, so-called Multi-agent sys- tems (MASs) for use in manufacturing gained much at- tention in the robotics and automation research community (Smith and Davis, 1981; Zlotkin and Rosen- schein, 1989; Simsarian and Mataric, 1995; Weiss and Sen, 1995). Due to their distributed nature, MASs promise, at least theoretically, some advantages that make them at- tractive structures for control and execution of manufac- turing processes. The main advantages are: (1) Modularity; (2) Robustness and fault tolerance; (3) Maintainability; (4) Extendibility. These features of MASs hold the potential of building manufacturing systems with greater ¯exibility than the currently used monolithic ones. In general, tasks that are to be accomplished by an MAS are assigned to one or more of its agents for execution. There exist a variety of methods regarding the process of task decomposition, negotiation, and assignment in MASs. One approach is to have the process controlled and/or managed by one dedicated agent called, e.g. a manager agent (Kaiser et al., 1996b). In other systems the burden of further decomposing a task, building a team for task exe- cution etc. is given to each agent that initially applied for a task and got it assigned (LaÈngle et al., 1995). In either case, knowledge about tasks, subtasks, and actions in general has to be transmitted between and interpreted by dierent agents. Especially in the case of applying learning methods to MASs in order to propagate task solutions from one agent to another, strong requirements are to be met by the representation structure used for action knowledge repre- sentation. For example, using the same identi®ers for ac- tions that have equivalent eects, and storing further symbolic information about the actions semantics etc. is necessary. These requirements will be analysed in detail in Section 4.1. When moving the focus of attention away from the MAS as a whole and directing it towards the agents themselves, Journal of Intelligent Manufacturing (1998) 9, 119±127 0956-5515 Ó 1998 Chapman & Hall