Inhalt des Dokuments
Room: MAR 5.065
Office hours: on request
Grasping is a peculiar topic in Robotics: on one hand we have physically motivated mechanical models which allow us to simulate complex manipulation in great detail. On the other hand, algorithms based on mechanical simulation aren't often applied to actual robot systems although these models have been around for a long time, sometimes for decades. Unfortunately, it seems, the approach of simulating on the mechanical level to find and optimize grasps seems to run into the curse of dimensionality, aggravated by the highly nonlinear behavior of collisions and contacts.
In my research I am exploring different routes of building and controlling robots. Seemingly paradoxical, increasing the complexity in the mechanics sometimes does not make the problem of grasping even more difficult, but actually simplifies it - adding the "right" mechanical complexity to an already complex interaction can make control simpler. I try to show in my research that it is possible to put this "simplifying complexity" directly into the morphology of the robotic hand. The robot's body does the heavy lifting in computation, hence we refer to this effect as "Morphological Computation".
The actual artefacts we build are so called Soft Hands. The hands are made from soft, deformable materials, provide compliant motion and complex, nonlinear deformations.
From the view of mechanics we push the frontiers of Soft Robotics, a young research field dedicated to understanding the advantages of completely soft robots. Integral to this research is the development of manufacturing methods of highly structured but soft artefacts, which are lagging far behind those of traditional rigid-bodied robotics in terms of accessibility and commercial availabillty.
From the view of Robotic Grasping we need to avoid formulating the grasping problem on a level above the mechanical description, to be able to increase morphological complexity to yield a simpler higher-level model. We investigate a possible representation using concepts such as Shape Adaptation and Environmental Constraints: Both describe patterns of interaction with real objects that are facilitated by compliant fingers and hands. Environmental Constraints are potential constraints to the free motion of the robot. Instead of viewing them as an obstacle to avoid we view them as helpful structures whose nonlinear interaction behavior (i.e. collision) can be used to increase certainty about the mechanical state of the world, such as the distance between hand and an object to grasp can be made more certain by first colliding with the surface the object rests on.
To bootstrap robots with novel grasping strategies, we also look at already proficient graspers and manipulators: adult humans. Humans can grasp even when they don't pay attention (i.e. dedicate mental resources) or when their senses are heavily impaired. These conditions resemble the state of the art of robot perception much better than unimpaired, attentive grasping, and our research indicates that humans not only apply different grasping strategies, but that humans seek collisions with (what we call) environmental constraints.
Besides these core research topics, I am also interested in the interaction of acting and sensing and the ramifications on the concept of representation.
- Soft Robotics
- Hand Grasping
- Sensorimotor Learning
- Sensorimotor Contingencies, Machine Learning, Bayesian statistics
2014 Best Student Paper Award at the Robotics: Science and Systems.
- 2012 - now PhD Student, Robotics and Biology Lab, TU Berlin
- 2008 - 2011 Middle European interdisciplinary Master in Cognitive Science, University of Vienna / Budapest University of Technology and Economics.
- 2004 - 2009 Bachelor of Science, Technische Informatik, Vienna University of Technology
- 2000 - 2005 Industrial Control Systems Developer, VA Tech