Motor Schemas in Robot Learning

Motor schemas used for robot learning are sequences of action that accomplish a goal-directed behavior, or a task. Motor schemas in robot learning are also known as movement primitives, basis behaviors, units of action, and macro actions. Rather than representing the simplest elementary actions available to the robot, such as a simple command to a robot actuator, schemas and motion primitives represent a higher-level abstraction of robot actions, such as “avoid obstacles”, “wander”, “walk”, “grasp a cup”, and “move to goal”. These schemas and motion primitives define control policies that are encoded with only a few parameters, and serve as the basis set, or movement vocabulary, of the robot. Such primitives are sufficient for generating the robot’s entire repertoire of motions via the combination of schemas or primitives. The schema can serve as a basis for robot learning, since it provides an abstraction that can be represented with fewer parameters, thus reducing the complexity of robot learning. This reduction in learning complexity allows robot learning to scale to more complex robots or tasks, thus making practical applications tractable.