Learning Predictive Knowledge to Optimize Robot Motor Control
Freek Stulp, Michael Beetz · 2008
Abstract — By observing the execution of their actions, cognitive robots become aware of their behavior. We describe a system that acquires such knowledge, and uses it to reflect future actions to autonomously avoid failures and optimize future actions. The system has three types of motor control knowledge, which are represented at different levels of abstraction, and acquired in different ways: 1) declarative knowledge to select actions, 2) procedural knowledge to execute actions, and 3) predictive knowledge to optimize action executions with respect to execution duration and success. The robots acquire predictive knowledge autonomously, by learning it from observed experience. The generality of the approach is demonstrated by applying the methods to three robotic platforms: a Pioneer soccer robot, a simulated articulated B21 in a kitchen environment, and a PowerCube arm. I.