Learning value functions with relational state representations for guiding task-and-motion planning

We propose a novel relational state representation and an action-value function learning algorithm that learns from planning experience for geometric task-and-motion planning (GTAMP) problems, in which the goal is to move several objects to regions in the presence of movable obstacles. The representation encodes information about which objects occlude the manipulation of other objects and is encoded using a small set of predicates. It supports efficient learning, using graph neural networks, of an action-value function that can be used to guide a GTAMP solver. Importantly, it enables learning from planning experience on simple problems and generalizing to more complex problems and even across substantially different geometric environments. We demonstrate the method in two challenging GTAMP domains.