@misc {PUBA282,
	title = {Tactile Exploration and Grasping Unknown Objects by Actively Learning Gaussian Process Implicit Shape Potentials},
	author = {Stanimir Dragiev AND Mark Toussaint AND Michael Gienger},
	year = {2013},
	month = {May},
	abstract = {Object estimation is an example where (a) promising
query points depend on observed ones, (b) acquisition of
observations is not for free {\textendash} neither w.r.t. time, nor energy
{\textendash} and (c) the purpose of estimation can range from obtaining
a very exact model to model just enough to grasp it. This
makes it an active learning scenario worth discussing. GPISP
is a probabilistic shape representation which is able to capture
the sensor uncertainty and expose it to the control algorithm.
The controllers we formulated on it in previous work can serve
as a basis in an active learning setting by introducing a bias
toward high-variance spots, or in explore-exploit policies. In
this paper, we interpret previous achievements in the light of
active and reinforcement learning, the relation of the control
laws to model improvement objectives, and discuss about graspspecific
properties. We argue that task specific utility terms can
augment the active learning objectives and make exploration
more task oriented. This is expected to solve a problem which
is more relevant in terms of sensing for grasping than the plain
object exploration is, in particular it should explore regions
which {\textendash} taken together {\textendash} are more suitable for grasping.},
	publisher = {MIT Press, Cambridge, Massachusetts},
	url = {Workshop of the 2013 Robotics Science and Systems Conference},
	booktitle = {Workshop of the Robotics Science and Systems Conference (RSS)},
	editor = {Nicholas Roy, Paul Newman and Siddhartha Srinivasa},
	city = {Berlin}
}