Eigenobjects: A Unified Framework for Robot-Focused 3D Object Representation
bcburch at cs.duke.edu
||Monday, April 25, 2016
||2:00pm - 4:00pm
||N311 North, Duke
One of the primary goals of robotics research is to enable robots to operate in unstructured
human-centric environments such as homes or offices. To enable this capability, robots will
need to reason about novel objects by generalizing from other previously encountered items.
Three main subtasks are required to facilitate this: object classification, pose estimation, and
3D estimation of unobserved portions of the object. We propose a PCA-based approach to representing
previously encountered (voxelized) 3D objects that learns low-dimensional spaces
that capture each object class well. Our preliminary data shows that such a representation
well suited to both classification and pose estimation, and also offers a natural extension to 3D
completion of novel objects. We show that given a novel and partially specified object, it is possible
to estimate the unobserved portions, estimate the object's pose, and assign it a class label.
We present initial results on several data-sets showing our method performing classification,
pose estimation, and object completion. As a next step, we propose extending our approach to
create hierarchical clusters of object classes, propagating information between object classes,
and learning parameterized functions (such as grasping controllers) that apply to portions of
the learned hierarchy. The result will be a single representation that allows information from
previously encountered objects to transfer to novel ones - even between classes - and supports
fast real-time queries with sub-linear time complexity in the number of object classes.
Advisor(s): George Konidaris
Committee: Carlo Tomasi, Katherine Heller, Kris Hauser