Efficient Multi-Step Planning for Robotic Manipulation

Multi-step planning is a promising technique for solving complex manipulation problems directly in the state space of the object, however, previous formulations were too slow to be practical except on problems with small search spaces. In this paper we propose and demonstrate two optimizations for this technique that together yield a significant speed-up on a previously used benchmark problem. The optimizations are 1. using the A* discrete search algorithm with proposed heuristics, and 2. a data structure for tracking the search graph. Depending on initial configuration, our results show a 2.7x-12.8x speed up in planning time as well as significant reductions in other search tree metrics. This faster planning algorithm also allows us to demonstrate the multi-step planning process on a combinatorially more complex tabletop planning problem.