Information Technology Reference
In-Depth Information
Action-Driven Perception for a Humanoid
Jens Kleesiek
1
,
3
, Stephanie Badde
2
, Stefan Wermter
3
, and Andreas K. Engel
1
1
Department of Neurophysiology and Pathophysiology,
University Medical Center Hamburg-Eppendorf, Hamburg, Germany
2
Department of Biological Psychology and Neuropsychology,
University of Hamburg, Hamburg, Germany
3
Department of Informatics, Knowledge Technology,
University of Hamburg, Hamburg, Germany
{
j.kleesiek,ak.engel
}
@uke.uni-hamburg.de,
Stephanie.Badde@uni-hamburg.de,
wermter@informatik.uni-hamburg.de
Abstract.
We present active object categorization experiments with a real hu-
manoid robot. For this purpose, the training algorithm of a recurrent neural net-
work with parametric bias has been extended with adaptive learning rates. This
modification leads to an increase in training speed. Using this new training al-
gorithm we conducted three experiments aiming at object categorization. While
holding different objects in its hand, the robot executes a motor sequence that
induces multi-modal sensory changes. During learning, these high-dimensional
perceptions are 'engraved' in the network. Simultaneously, low-dimensional PB
values emerge unsupervised. The geometrical relation of these PB vectors can
then be exploited to infer relations between the original high dimensional time
series characterizing different objects. Even sensations belonging to unknown ob-
jects can be discriminated from known (learned) ones and kept apart from each
other reliably. Additionally, we show that the network tolerates noisy sensory
signals very well.
Keywords:
Active Perception, RNNPB, Humanoid Robot.
1
Introduction
Motor actions determine the sensory information that agents receive from their envi-
ronment. Combining sensory and motor processes dynamically facilitates many tasks,
one of those being object classification.
The intention of this experiment is to provide a neuroscientifically and philosophi-
cally inspired model for
what do objects feel like?
For this purpose, we stress the active
nature of perception within and across modalities. According to the sensorimotor con-
tingencies theory [1], actions are fundamental for perception and help to distinguish the
qualities of sensory experiences in different sensory channels (e.g. 'seeing' or 'touch-
ing'). O'Regan and Noe actually suggest that “seeing is a way of acting” [1]. Exactly
this is mimicked in our computational study.
It has been shown that if the fruit fly drosophila cannot recognize a pattern it starts
to move [2]. It is also known that flies use motion to visually determine the depth of
