As we have seen, the ALLIANCE and L-ALLIANCE architectures allow robot teams to accomplish missions of loosely coupled, largely independent subtasks with a sig-nicant degree of coherence. However, the extent of coherence attainable by these teams has been shown to be dependent upon the knowledge individual robots possess concerning the current actions and previous performance of their teammates. This knowledge can actually be viewed as partial global information about the current state and intentions of the robot team. The more limited this global knowledge be-comes, the more each robot depends upon its own local knowledge for action selection, which may in turn decrease the coherence of the team. However, due to the design of ALLIANCE and L-ALLIANCE, the use of the global knowledge is fortunately not detrimental to the processing requirements of the individual robots. Thus, the use of global knowledge can be incorporated into ALLIANCE and L-ALLIANCE without impacting the team performance.
It is interesting to step back for a moment and consider whether this principle of
\increased global knowledge implies increased coherence" holds for a dierent type of cooperative robot mission | namely, one requiring spatial coordination among robot team members. It is appealing to be able to develop control laws that utilize strictly local information such that the desired group coordination emerges from the interaction of the local control laws. Indeed, research has shown that certain types of spatial coordination missions can be achieved using local control knowledge alone [Deneubourget al., 1992, Drogoul and Ferber, 1992, Franklin and Harmon, 1987, Goss and Deneubourg, 1992, Kube and Zhang, 1992, Miller, 1990, Steels, 1990, Stilwell and Bay, 1993, Theraulazet al., 1990]. However, the question that remains is determining the degree to which group coordination and coherence can be achieved for a given application with purely local control knowledge, and when more global knowledge is needed to obtain the desired results. While I do not attempt to answer this question thoroughly here, I do describe the results of one case study | keeping formation |
179
which, at rst glance, appears to be an application that can be solved using local control alone. However, upon further investigation, we discover that local control rules are not sucient to obtain the desired level of performance for the mission.
The following sections rst distinguish between the notions of global control and local control and then examine the tradeos between the two types of control laws.
I present the \Keeping Formation" case study which stimulated my thoughts on the local versus global control issues, discussing the design and implementation of several alternative control strategies and the results. This chapter concludes with a summary of the general principles and guidelines derived through this case study. (See [Parker, 1993a] for a related discussion of this issue.)
7.1 Descriptions of Global and Local Control
In practice a continuum exists between strictly global and strictly local control laws.
Thus, the control laws guiding a robot will probably use a mixture of local and global knowledge, rather than adhering strictly to one type alone. To simplify the discussion, however, these types are considered separately in this section, which compares and contrasts these two types of control.
7.1.1 Global Control
Global control laws utilize the global goals of the cooperative team and/or global knowledge about the team's current or upcoming actions to direct an individual robot's actions. With these laws, a robot is able to inuence its own actions toward team-level goals that cannot be sensed in its own local world. To better understand the implications of the use of global control laws, let us look individually at the two types of information utilized by these laws: global goals and global knowledge. The global goals of a team indicate the overall mission that the team is required to accomplish. These goals are typically imposed upon the team by a centralized controller, such as a human or another autonomous robot. Often this controller is a robot from outside the cooperative team rather than from within, although it is not uncommon to have a leading robot within the team specifying these goals.
Of particular impact on the design of cooperative teams is the time at which the global goals become known[Payton, 1991]. If the goals are known and xed at design-time, then it may be possible to incorporate these goals implicitly into the control laws of each robot. Whether this can be done depends on the proper match between the sensing capabilities of the robots and the sensing requirements of the global goals.
If all the information required for a robot to act consistently with the global goals can be sensed locally by that robot at run-time, then the global goals can be designed
7.1. DESCRIPTIONSOF GLOBAL AND LOCAL CONTROL 181 into the robot. On the other hand, if the goals are not xed or known at design-time, then they obviously cannot be designed into the robots. In this case, the robots must possess the capability to obtain and appropriately act upon the goals provided at run-time.
The second type of information used by global control laws, global knowledge, refers to the additional information that may be necessary for the cooperative team to achieve the global goals. This information typically indicates what other robots in the team are doing or are going to do, or what the environment looks like in relation to the current cooperative task. By denition, all such information is normally not avail-able to the individual robots through their sensors (other than their communication channels); if it were, then I would consider it to be local information.
How does a robot obtain this global knowledge? Several methods are possible.
Perhaps the most obvious manner is for a centralized informant (either a human or an autonomous robot either inside or outside of the robot team) to explicitly communicate the information directly to the team as it becomes available. The robots can then utilize this explicitly communicatedinformation as advice, along with locally sensed data, to undertake appropriate actions which are consistent with the global goals. A second method of obtaining global knowledge, albeit in an approximate form, is for robots to passively observe and interpret the actions of another robot as described in the earlier chapter on action recognition. Combined with some goal recognition, this method would allow a robot not only to interpret a teammate's current actions, but also to predict that robot's future actions. In a sense, this method utilizes implicit communication, since the observing robot receives information from the actions of the observed robot.
The use of global goals and information is not without its shortcomings, however.
Adequate global information may not be available to achieve the desired global goal.
Even with global knowledge, a robot may still not exhibit optimal global behavior un-less it utilizes all of the global knowledge available. Processing this global information requires time and resources, both of which are usually limited in real-world applica-tions. If the global goals or information is changing often enough, the robot may not be able to act upon the global knowledge before it becomes out-of-date. Indeed, in some situations, global control of any kind will be impossible, thus mandating the use of local control.
7.1.2 Local Control
Localcontrol laws, on the other hand, guide a robot's actions based on the proximate environment of that robot. Such information is derived from the robot's sensory capabilities, and thus reects the state of the world near the robot. Local control laws allow robots to react to dynamic changes in their environment without relying
on preconceived plans or expectations of the world. As I have noted, careful design of the control laws can allow global functionality to emerge from the interaction of the local control laws of the individual robots. For example, Franklin and Harmon [Franklin and Harmon, 1987] have shown that a global cooperative hunting behavior emerges from the use of three local cooperative control laws: cooperative pursuit, triangulation, and encirclement. These control laws are appealing because of their simplicity and power to generate globally emergent functionality.
However, local control laws also have their limitations | certain global goals cannot be attained through the use of local control laws alone. In some cases, it may be possible to utilize local control laws to achieve an approximation to the optimal results, which may be totally acceptable for many applications. However, since local control relies strictly on features of the environment that can be sensed, those aspects of global goals that have no physical manifestation in the world cannot be acted upon by local control laws.
7.2 Keeping Formation Case Study
Let us now look at the keeping formation case study to see what we can learn about the level of control attainable with various combinations of local and global control.
I have implemented and evaluated several control strategies along the local versus global spectrum by performing a wide range of experiments in simulation. For each of the control strategies, I measured the results quantitatively by collecting data on the mission completion time and amount of robot error in performing the mission. This section describes these results, rst dening the mission performed by the robots, briey reviewing the related work in this area, and then discussing the results of experiments with four control strategies that vary in the amount of global and local information.
7.2.1 Task Description
The keep formation task requires a group of robots to stay in formation with one another (i.e. remain aligned side to side) while the leader of the group follows a prespecied route and while all robots avoid obstacles as they appear (see gure 7-1).
Each of these robots has the ability to sense the location of its neighboring robots relative to itself (local knowledge).
The global goal of this task is twofold: rst, the robots should reach their destina-tion as quickly as possible, and, second, they must maintain the specied formadestina-tion in a manner that appears to a casual human observer to be human-driven, meaning that the robots should not allow huge or \unnatural" (an admittedly subjective measure)