| Summary: | Planning in robotics environments is difficult in part due to continuous state and action spaces. One approach to this challenge is to use bilevel planning, where decisionmaking occurs in multiple levels of abstraction. However, the efficacy and efficiency of bilevel planning relies on the underlying set of state and action abstractions. It is impractical to assume these abstractions as given (i.e. hand-designed by humans), so instead the agent should learn them, for example by exploring and interacting with its environment under the guidance of a teacher, from whom the robot may query expert knowledge. This is more difficult than the typical active learning problem setting because the robot must take actions to get to a state before it can make useful queries of the teacher about that state. This work develops an active learning framework for learning state abstractions (predicates) for effective and efficient task and motion planning. Given the names and arguments of the predicates in an environment and very few pre-labeled examples, the agent is able to learn predicate classifiers that enable it to successfully complete test tasks.
|
|---|