Robot Learning Constrained by Planning and Reasoning.

摘要: Robot learning is usually done by trial-anderror or example. Neither of these methods takes advantage prior knowledge or of any ability to reason about actions. We describe two systems. In the first, we learn a model robot's This used in simulation search for sequence actions that achieves goal traversing rough terrain. Further used compress results this into set situation-action rules. second system, we assume robot has some effects and can use plan The qualitative states result from are as constraints trial-and-error learning. approach greatly reduces number trials required learner. method demonstrated on problem bipedal walk.