LEARNING BASED PLANNING FOR HIGH LEVEL GOALS

Abstract

Given a con guration involving n objects in an environment, the planning problem considered in this theses is to nd a plan that re- arranges these objects so as to place a new object. The challenging aspect here involves deciding when an object can be placed on top of another object. Here only de ning standard planning operators would not su ce. For instance, using these operators we can come up with actions that may be performed at a state but it should not be performed. So we introduce the notion of safe actions whose outcomes are consistent with the laws of physics, commonsense, and common practice. A safe action can be performed if a robot perform- ing the action knows the knowledge of the situation. We developed a knowledge engine using a supervised learning technique. However, unlike the common task of learning functions, our approach is to learn predicates|that evaluate to binary values. By learning such a predicate a robot would be able to decide whether or not an object A can be placed on top of another object B. Once the robot learns the predicate, planning can be carried out and the plans contain only safe actions. We also suggest a method to handle objects previously not seen in the training set.