Guided Sequential Manipulation Planning Using a Hierarchical Policy

We introduce a hierarchical policy structure that selects high-level actions for effective task and motion planning (TAMP) in sequential manipulation tasks. For such problems, scalability of the methods is a major challenge, due to the combinatorial complexity of possible discrete decisions. To overcome this, we propose to learn an upper-level policy that selects the next manipulation action, and a lower-level policy that decides on the end-effector and objects to be involved in the action given the encoded current state. We demonstrate the generalizability of our approach in various pick-and-place experiments. We further show that the time and space complexity is significantly reduced compared to a state-of-the-art TAMP framework especially for tasks involving many objects.