A two-level approach for intricate manipulation planning

In the past few years several research efforts have aimed to produce robots capable of operating in environments inhabited by humans. This requirement places some tough constraints on a robot; among them, the ability to solve manipulation tasks that come as second nature to people. Manipulation problems are difficult because they present narrow passages (often, the kinds of possible or useful maneuvers are included in a very small subset of all possible maneuvers) as well as very high dimensional configuration spaces. Manipulation tasks often require both arms to be used, and the grasp on a manipulated object matters. Furthermore, the grasp may need to be changed along the execution of the task. To make planning for such problems tractable, we propose a two-level planning architecture inspired by the way humans tackle such problems: plan a motion for the object first, then the sequence of manipulation actions. This approach can alleviate both the narrow passage and the high-dimensionality problem, as well as offer the possibility for human assisted motion planning. We offer an intuitive justification for the method and validate it experimentally.