Selected Work on Robotic Grasping and Manipulation

Topological Ideas for Grasp Synthesis

I am interested in novel topologically motivated representations that can be used for robotic grasping. In our recent work "Grasping Objects with Holes: A Topological Approach" (ICRA 2013, with J. A. Stork, D. Kragic), we investigate using topological features of an object with holes. For this purpose, we make use of the first homology group $H_1(X)$ of a surface $X\subset \mathbb{R}^3$. $H_1(X)$, when computed with coefficients in a field, is a vector space whose non-zero elements correspond to equivalence classes of non-contractible loops on $X$. Our work introduces these topological notions for grasping. We adapt ideas coming from the Gauss linking integral and the notion of winding numbers in topology to define a simple controller leading a robotic hand to securely clasp around such loops. Recently, we combined this approach with motion planning and introduced a notion of virtual linking. Our paper A Topology-based Object Representation for Clasping, Latching and Hooking" (Humanoids 2013, J. A. Stork, F. T. Pokorny, D. Kragic) has been selected as runner-up for the best conference award at Humanoids 2013. Our work in Danica Kragic's lab is partially funded by the EU FP7 project TOMSY which investigates topological techniques for dexterous manipulation.

Grasp Moduli Spaces and Continuous Representations

Our recent work "Grasp Moduli Spaces" (RSS 2013, with K. Hang, D. Kragic) is part of our efforts to introduce new continuous parametrizations to robotics which are capable of describing shapes and grasps in a single space. Based on such spaces we are interested in developing deformation based approaches to manipulation which are capable to transfer grasps across similar object instances.
Our current work in this direction focusses on developing novel such spaces which can easily be used with real world sensor data.

Grasp Quality Functions

Grasp synthesis algorithms commonly use grasp quality assessment functions which measure how physically stable or suitable a grasp is. In our work Classical Grasp Quality Evaluation: New Theory and Algorithms (IROS 2013, with D.Kragic), we investigate new efficient methods for computing a popular grasp quality measure introduced by Ferrari and Canny. The computation of this grasp quality measure classically relies on an approximation of friction cones by polyhedral cones. We make precise error bounds when applying this approach and prove that the underlying grasp quality measure $Q$ is in fact Lipschitz continuous with a particular Lipschitz constant. This implies a novel analytic bound on the variability of grasp quality around any given grasp which is useful in the presence of noise, or when a deformation based grasp synthesis approach is to be considered. Another aspect of grasp evaluation I am interested in is the development of task specific grasp quality measures.