A General Manipulation Task Planner

This paper addresses the manipulation planning problem which deals with motion planning for robots manipulating movable objects among static obstacles. We propose a general manipulation planning approach capable to deal with continuous sets for modeling both the possible grasps and the stable placements of the movable object, rather than discrete sets generally assumed by the existing planners. The algorithm relies on a topological property that characterizes the existence of solutions in the subspace of configurations where the robot grasps the object placed at a stable position. This property leads to reduce the problem by structuring the search-space. It allows us to devise a manipulation planner that directly captures in a probabilistic roadmap the connectivity of sub-dimensional manifolds of the composite configuration space. Experiments conducted with the planner demonstrate its efficacy to solve complex manipulation problems. Several applications such as robot programming, manufacturing or animation of digital actors require some real or virtual artifact to move and manipulate objects within an environment composed of obstacles. Manipulation planning concerns the automatic generation of the sequence of robot motions allowing to manipulate movable objects among obstacles. The presence of movable objects, i.e. objects that can only move when grasped by a robot, leads to a more general and computationally complex version of the classical motion planning problem [14]. Indeed, the robot has the ability to modify the structure of its configuration space depending on how the movable object is grasped and where it is released in the environment. Motion planning in this context appears as a constrained instance of the coordinated motion planning problem. Indeed, movable objects can not move by themselves; either they are transported by robots or they must rest at some stable placement. The solution of a manipulation planning problem consists in a sequence of sub-paths satisfying these motion restrictions. In related literature (e.g. [2,14]), motions of the robot holding the object at a fixed grasp are called transfer paths, and motions of the robot while the object stays at a stable placement are called transit paths. Consider the manipulation planning example illustrated by Figure 1. The manipulator arm has to get a movable object (the bar) out of the cage, and to place it on the other side of the environment. Solving this problem requires to automatically produce the sequence of transfer/transit paths separated by grasps/ungrasps operations, allowing to get one extremity of the bar out of the cage; the manipulator can then re-grasp the object by the extremity that was made accessible by the previous motions, perform a transfer path to ex2 Thierry Siméon et al. Fig. 1. A manipulation planning problem (left) and its computed solution (right) tract the bar from the cage, and finally reach the specified goal position. This example shows that a manipulation task possibly leads to a complex sequence of motions including several re-grasping operations. A challenging aspect of manipulation planning is to consider the automatic task decomposition into such elementary collisions-free motions. Most of existing algorithms (e.g. [1,2,5,12,18]) assume that finite sets of stable placements and possible grasps of the movable object are given in the definition of the problem. Consequently, a part of the task decomposition is thus resolved by the user since the initial knowledge provided with these finite sets has to contain the grasps and the intermediate placements required to solve the problem. Referring back to the example, getting the bar out of the cage would require a large number of precise grasps and placements to be given as input data. In this paper, we describe a more general approach based onto recent results presented in [22,23]. The proposed approach is capable to deal with a continuous setting of the manipulation problem. It allows us to design a manipulation planner that automatically generates among continuous sets the grasps and the intermediate placements required to solve complicated manipulation problems like the one illustrated onto Figure 1. The approach relies on a topological property first established in [3] and recalled in Section 1. This property allows us to reduce the problem by characterizing the existence of a A General Manipulation Task Planner 3 solution in the lower dimensional subspace of configurations where the robot grasps the movable object placed at a stable position. Section 2 describes the proposed two-stage approach. First, Section 2.1 shows how the connected components of this sub-space can be captured in a probabilistic roadmap computed for a virtual closed-chain system. Using the visibility technique [20] extended to deal with such closed systems [7], this first stage of the approach captures the connectivity of the search space into a small roadmap generally composed of a low number of connected components. Connections between these components using transit or transfer motions are then computed in a second stage (Section 2.2) by solving a limited number of point-to-point path planning problems. The details of an implemented planner interleaving both stages in an efficient way are described in Section 3. Finally, we discuss in Section 4 some experiments and the performance of the planner.