Learning the search heuristic for combined task and motion planning

Autonomous robots have to plan two intricately dependent levels: task and motion. One promising approach is to plan task and motion simultaneously, yielding a sequence of high level actions that is guaranteed to have valid motion plans. In this paper, we present our work on such planning system whose backbone is the ability to estimate the cost of action sequences. This cost essentially encodes information about motion feasibility and optimality criteria. Concretely, the cost prediction serves as the heuristic for search over a task motion multigraph. The experiment results show that the proposed approach makes the planning progressively more efficient as well as ε-optimal. It means that the wasted computations are more and more reduced over planning attempts and that the complete plans found are guaranteed to have costs no more than a factor of (1 +ε) greater than the optimal. This suggests that the heuristic along with its learning formulation are justifiable and that the designed feature vector is sufficient for learning. In addition, we found that online learning during search offers better utility than the offline.