Machine Learning and Friends Lunch
Generalization and Differentiation of Robot Manipulation Skills
Robert Platt
UMass
Abstract
One of the hallmarks of natural intelligence is the ability to robustly generalize control knowledge to a variety of situations. However, many robot systems cannot apply a skill learned in one context to a different context. Since robots that act in open environments almost never encounter exactly the same situation twice, an inability to generalize prevents a robot from re-using control knowledge. In addition to enabling a control policy to be used in more situations, re-using control knowledge is important because it reduces the underlying high-dimensional robot control problem to the easier problem of combining a set of generalized skills.
This talk proposes an action schema that represents a generalized instance of a robot control task. The action schema encodes a task as a sequence of objectives where each objective can be met by a member of a corresponding set of low-level control choices. By selecting different instantiations of the sequence of task objectives, the action schema can describe a variety of related concrete instances of behaviors. However, some of these schema instantiations may result in behaviors that do not accomplish schema objectives. In addition, the correct instantiation of a schema may depend on physical variables in the external world that are unknown to the robot. I describe an on-line algorithm that prunes those instantiations that are not correct and allows the system to take control actions that reveal contextual information relevant to the schema's subsequence decisions.