Foundations of Robotics
Seminar, February 4, 2009
Time
and Place | Seminar Abstract
Robust Object Shape Models for Grasping
Jared Glover
Department of Electrical Engineering and Computer Science
MIT
NSH 1507
Talk 4:30 pm
Robot manipulators typically rely on complete
knowledge of object geometry in order to plan motions and
compute grasps. However, manipulating real-world objects is
extremely challenging, since the shape of new instances of known
objects classes may vary from learned models, and since deformable
objects may appear in new configurations that do not match previous
observations.
In this talk I will describe an algorithm for learning generative
models of object geometry for the purposes of manipulation;
these models capture both non-rigid deformations and object
variability within a known class. The models can be used to
recognize objects in a single image based on the visible portions
of each object contour. The complete geometry of the object can
then be estimated to allow grasp planning. To our knowledge,
our work is one of the first to perform probabilistic inference
of deformable objects from partially occluded views. We show
examples of learned models from image data and demonstrate
how the learned models can be used by a manipulation planner
to grasp objects in cluttered visual scenes.
Jared Glover received a B.S. in Computer Science from Carnegie Mellon in 2005
and a Master's in Electrical Engineering & Computer Science from MIT in 2008. While at Carnegie Mellon, he worked extensively on the Nursebot project, where
he developed software for a robotic walker and for Pearl, a humanoid robot.
He currently works at Two Sigma Investments in New York.
The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.