Learning prospective pick and place behavior

摘要

When interacting with an object, the possible choices of grasping and manipulation operations are often limited by pick-and-place constraints. Traditional planning methods are analytical in nature and require geometric models of parts, fixtures and motions to identify and avoid the constraints. These methods can easily become computationally expensive and are often brittle under model or sensory uncertainty. In contrast, infants do not construct complete models of the objects that they manipulate, but instead appear to incrementally construct models based on interaction with the objects themselves. We propose that robotic pick-and-place operations can be formulated as prospective behavior and that an intelligent agent can use interaction with the environment to learn strategies which accommodate the constraints based on expected future success. We present experiments demonstrating this technique, and compare the strategies utilized by the agent to the behaviors observed in young children when presented with a similar task.