Learning operators for manipulation planning

摘要

We describe a method for learning planning operators for manipulation tasks from hand-written programs to provide a high-level command interface to a robot manipulator that allows tasks to be specified simply as goals. This is made challenging by the fact that a manipulator is a hybrid system-any model of it consists of discrete variables such as “holding cup” and continuous variables such as the poses of objects and position of the robot. The approach relies on three novel techniques: the action learning from annotated code uses simulation to find PDDL action models corresponding to code fragments. To provide the geometric information needed we use supervised learning to produce a mapping from geometric to symbolic state. The mapping can also be used in reverse to produce a geometric state that makes a set of predicates true, thus allowing desired object positions to be generated during planning. Finally, during execution of the plan we use a partially observable Markov decision problem-based planner to repair the initial plan when unforeseen geometric constraints prevent actions from being executed.