Author: David Andreae
Source: GZipped PostScript (63kb); Adobe PDF (241kb)
We are developing a system called GRAM that can learn structural concepts from examples in the domain of two-dimensional physical objects. The components of the system are intended to be eventually incorporated into a general purpose instructable robot. gram is given information from a vision system about an observed object, from which it constructs a hierarchically organised part-based description. This is matched with an existing concept description, and the resulting correspondences between the parts are used to generalise the concept appropriately.
This paper gives a brief overview of the representation scheme and the matching algorithm.