Page 321 - Introduction to AI Robotics

304
8.5 Goals 8 Multi-agents
The final dimension for characterizing a collection of multi-agents is how the
robot works on a goal. If all the robots in the collection work on attaining the
same explicit goal, then they are said to share a single goal, versus having
individual goals.
An example of robots working a single goal is the winning team for the
Office Navigation event in the 1996 AAAI Mobile Robot Competition. 78 The
office navigation event had a robot that was supposed to search a series of
rooms, find an empty conference room, and then go to a list of rooms where
people were and tell them that a meeting was going to begin in the empty
conference room. The event was originally conceptualized as a single agent
task, but the SRI entry under the direction of Kurt Konolige consisted of three
robots. 62 Each of the three robots ran the Saphira architecture (see Ch. 7) and
were coordinated by a central workstation. While the robots were responsi-
ble for autonomous navigation, their goals were set by the central strategy
agent. Even though they were navigating through different parts of the office
maze, the robots were working on a single goal and the software agents on
the central workstation were explicitly coordinating the actions of the robots.
The robots were able to find an empty room and inform the attendees in 4
minutes and 30 seconds. The next best time was close to 10 minutes.
An example of purely reactive robots working on individual goals is a
9
problem originally posed by Ron Arkin: a group of robotic space “ants”
foraging for mineral rich asteroids or Near Earth Objects (NEOs). If each
robot in the group forages for its own asteroid, then they have individual
goals. (Notice that a behavior that permits them to notice other robots and
be repulsed will help disperse the robots.) If the robots are programmed so
that they will all go to one specific asteroid, then they share a common goal.
Emergent cooperation is not the same thing as having a single goal. For
example, suppose the robotic space ants are programmed to go to the near-
est non-moving asteroid and bring it back to base. Each robot might have a
set of behaviors: find-stationary-asteroid, move-to-asteroid, push-asteroid-
to-home, and avoid-robots. The find-stationary-asteroid could be done with
a random potential field (in 3 dimensions, of course). An attractive “asteroid-
tropic” potential field could be used for the move-to-asteroid behavior. Like-
wise an attractive field could be used for the push-asteroid-to-home behav-
ior, where the robot tries to stay behind the asteroid as it moves to home
rather than avoid the asteroid. Avoid-robot could be done with a repul-
sive field. These behaviors give the robots individual goals, since there is