Solving Task Space Problems with Multi-Step Planning

Modern robotics research has developed a mature family of planners for solving robot motion planning problems, but task space problems (where we want to reason about objects) remain difficult. Other approaches for solving what are often referred to as combined task and motion planning problems rely on bringing logical problem structure into the design of the high level planner. While programming assumptions around logical problem structure cannot be avoided, we propose the multistep planning approach which maintains high level generality while pushing these logical assumptions to a lower level of the program. Through experiments on a real robot and in simulation we demonstrate the ability of this architecture to solve real (though small) problems. We also demonstrate how carefully chosen heuristics can be key to making this approach faster. Note to Practitioners: This paper proposes techniques for using multi-step planning to solve problems defined in the way we would like objects to move in the space, rather than how we would like the robot to move. Successful applications of this technique will depend of the ability of the practitioner to define useful families of planning spaces, and provide algorithms that can plan and sample states in those families. Additionally, in order to get good performance, particularly on larger problems, the practitioner should provide an informative heuristic that matches the cost function used in the planning process. Our results suggest that with good heuristics, multi-step planning could be a practical technique for problems with relatively small search complexity, and suggest the potential for developing stronger heuristics to target larger problems.