Basis Behaviors

Behavior-based control is one of the three dominant paradigms for robot control. For more details please look here.

In 1991 I introduced the notion of basis behaviors, a means for facilitating principled behavior synthesis and analysis in behavior-based systems. The basis behavior set of a system provides elements that are not further reducible to each other and that, when composed by sequential or concurrent execution, produce the complete behavior repertoire for the system. Such basis behaviors are constructed, learned, or evolved from stable, robust interaction dynamics between the agent/robot and its environment, and serve as a substrate for the system's more complex behaviors.

Biology provides evidence in support of basis behaviors at a variety of levels. A particularly clean and compelling case can be found in motor control. A relatively small set of basis vector fields, found in the frog and rat spine, generate the animal's entire motor behavior repertoire from appropriate combinations of the basis vectors. The described motor basis behaviors are a result of two types of constraint optimization: the dynamics of the manipulator and the dynamics of the motor tasks. In the case of motor control, the behaviors constitute prototypical reaches, grasps, throws, strides, and so on.

My work generalized the notion of basis behaviors to multi-robot interactions, and demonstrated how a small set of basis behaviors per robot can be used to demonstrate a rich repertoire of individual and group-level behaviors, including following, flocking, homing, herding, aggregation, dispersion, and formations. My PhD thesis outlined desirable criteria for selecting and evaluating basis behaviors, and subsequent work addressed learning and evolution of such basis sets.

We have successfully applied basis behaviors to single and multi-robot control, as well as single and multi-robot learning of various tasks. To see an example of basis behaviors in the multi-robot domain, look here for a 2.5 minute 15.2MB MPEG video of our robots performing avoidance, following, aggregation, dispersion, homing, and some compound behaviors (flocking and foraging). To see examples of articulated control using basis behaviors, look here. Most recently we are exploring the role of basis behaviors as visual-motor primitives in learning by imitation and looking at data-driven automated extraction of primitives. We are also extending the notion of behaviors to be used as more abstract representation. The list of papers below highlights some of this work. For details on our current related research efforts, please see my projects page and my publications page.

The following is a selection of my papers specifically on the topic of basis behaviors: