ROBOTICS AND AUTOMATION MAGAZINE - SPECIAL ISSUE ON MOBILE MANIPULATION 1 Simultaneous Grasp and Motion Planning Nikolaus Vahrenkamp, Member, IEEE, Tamim Asfour, Member, IEEE, and R¨ udiger Dillmann, Fellow, IEEE Abstract—In this work, we present an integrated approach for planning collision-free grasping motions. The proposed Grasp–RRT planner combines the three main tasks needed for grasping an object: building a feasible grasp, solving the inverse kinematics problem and searching a collision-free trajectory that brings the hand to the grasping pose. Therefore, RRT-based algorithms are used to build a tree of reachable and collision- free configurations. During RRT-generation, grasp hypotheses are generated and approach movements toward them are computed. The quality of reachable grasping poses is evaluated via grasp wrench space analysis. We also present an extension to a dual arm planner which generates bimanual grasps together with corresponding dual arm grasping motions. The algorithms are evaluated with different setups in simulation and on the humanoid robot ARMAR-III. Index Terms—Grasp Planning, Motion Planning, Humanoid Robots. I. INTRODUCTION H UMANOID robots are designed to assist people in daily life and to work in human-centered environments. In contrast to industrial applications, where the environment is structured to the needs of the robot, humanoids must be able to operate autonomously in non-artificial surroundings. One essential ability for working autonomously is to grasp a completely known object for which an internal represen- tation is stored in a database (e.g. information about shape, weight, associated actions or feasible grasps). Furthermore, the robot should be able to grasp objects for which the internal representation is incomplete due to inaccurate perception or uncertainties resulting in an incomplete knowledge base. The task of grasping an object induces several subtasks that have to be solved, like searching a feasible grasping pose, solving the inverse kinematics (IK) problem and finding a collision-free grasping trajectory. With the algorithms pro- posed in this article all these problems are solved with one probabilistic planning approach based on Rapidly Exploring Random Trees (RRT) [1]. The Grasp–RRT planning algorithm combines the creation of feasible and reachable grasps with the search for collision- free motions and thus no pre-calculated grasping data is needed. This online search for feasible grasping configurations has the advantage that the search is not limited to a potentially incomplete set of offline generated grasps. Furthermore, online generated requirements or constraints, that have to be met by the the grasping configuration, can be implicitly considered. Since such constraints (e.g. don’t grasp at a specific part of the object, consider post-grasping stability for transportation, etc) usually limit the number of feasible grasps that can be applied, N. Vahrenkamp, T. Asfour and R. Dillmann are with the Institute of Anthropomatics, Karlsruhe Institute of Technology (KIT), Germany, e-mail: (vahrenkamp,asfour,dillmann)@kit.edu. Manuscript received October 03, 2011; revised February 12, 2012. Fig. 1. A bimanual grasping trajectory. the object’s grasp database must be huge in case offline generated data is used. By generating the grasp hypotheses online, the Grasp–RRT approach avoids a filter step on offline generated discretized grasping data. Further, the search for a feasible grasp is focused on reachable configurations and thus the computation of grasping poses is only performed for positions that can be reached by the robot. This focus on the reachable part of the object allows an efficient implementation, as demonstrated by the experiments. The algorithms can be applied for single and dual arm planning problems and even when just a rough estimation of an unknown object is given, an approximated 3D model can be used to search grasping poses online. The proposed Grasp–RRT planner was originally presented in [2], where we used an efficient grasp measurement based on a low dimensional space related to impacting forces resulting from grasping contacts. In this work we utilize the more common approach of analyzing the Grasp-Wrench space for determining the quality of a grasping hypotheses. Additionally we present a more comprehensive evaluation including the growth in grasp quality over time and a comparative study that compares the algorithm to classical approaches. In the next section, related work dealing with planning motions for grasping is presented. The three parts of the Grasp–RRT algorithm (computing grasping poses, generat- ing approach movements and the online grasp quality mea- surement) are discussed in Section III. In Section IV the Bimanual Grasp–RRT algorithm, an approach for generating dual arm grasping motions, is presented. Several experiments for planning single arm and bimanual grasping motions in simulation and on the humanoid robot ARMAR-III (see Fig. 1) are discussed in Section V.