BRIAN

BRIAN (Brian Reacts by Inferential ActioNs) is an effective system to define and manage the combination of control modules, each implementing a behavior. It has been designed to be modular and fast. Presently, it implements an approach described in the papers mentioned below.

To each behavior is associated a set of fuzzy predicates that state when it is appropriate to apply it; for instance, it has no sense to Kick-the-ball if the robot is not controlling the ball. These CANDO conditions are evaluated to select the behaviors that can trigger at a given moment.

Another set of fuzzy predicates is associated to each behavior, and are used to give a weight to the action it proposes. These WANT conditions include context predicates (e.g., it is better if the robot kicks the ball in a direction where there are no opponents), social predicates (e.g., the robot should not Go-to-ball if a teammate is committed to do so), and goal predicates (e.g., if the robot has to score goals, it would be fine if it kicks the ball to the opponent goal).

In the present implementation, CANDO conditions are used to build clusters of compatible behaviors, and WANT conditions to weight the actions proposed by the different behavior modules that are active in a given situation. Presently, weighted actions are composed together, but winner-take-all selection can also be implemented.

We have also implemented the possibility to have hierarchies of behavior modules, so that the action proposed by a module is an input for a higher level module that can decide to filter it. This makes it possible to implement, for instance, an AvoidObstacle behavior that avoids obstacles while taking into account the actions that had been proposed and had contributed to the final actions if the obstacle was not present.

We are presently using BRIAN to control either Robocup robots, which you can see in action here and service robots which you can see here In the first case, also the single control modules are implemented by fuzzy rules, in the second, some of them are simple algorithms.

Principal investigators
A. Bonarini, M. Matteucci, M. Restelli

Research contributors
F. Barna, G. Invernizzi, T.H. Labella

Related Papers