Learning Perception-Action Cycles in a Driving School Scenario (DRIVSCO)
Funding Agency: CEC
Funding period: 2006-2009
URL: www.pspc.dibe.unige.it\~drivsco
Contact Person
Silvio P. Sabatini
Department of Biophysical and Electronic Engineering
University of Genoa
Via Opera Pia, 11a
I-16145 Genova, ITALY
tel: (+39) 10 353 2092
fax: (+39) 10 353 2777
e-mail: silvio.sabatini AT unige.it
Participants
Coordinator:
Universitaett Goettingen (DE)
Partners:
Westfaeliche Wilhems-University Muenster (DE)
University of Southern Denmark, Odense (DK)
Katholieke Universiteit Leuven (BE)
University of Genova (IT)
University of Granada (ES)
Vytautas Magnus University, Kaunas (LT)
HELLA Hueck KG & Co, Lippstadt (DE)
Abstract
Most technical systems, for example cars, must work reliably at key-turn. Therefore, such systems almost always employ conventional control strategies. Biological systems, on the other hand, learn. In the beginning they are functional only at a very basic level from which they improve their skills. No-one would, however, want to use a learning car, which could in the beginning barely steer. Thus, learning techniques have not really entered turn-key applications so far. The goal of DRIVSCO is to devise, test and implement a strategy of how to combine adaptive learning mechanisms with conventional control, starting with a fully operational human-machine interfaced control system and arriving at a strongly improved, largely autonomous system after learning, that will act in a proactive way using different predictive mechanisms.
DRIVSCO seeks to employ closed loop perception-action learning & control to cars and their drivers; combining for the first time advanced (largely hardware based) visual scene analysis techniques with supervised sequence learning mechanisms into a semi-autonomous and adaptive control system for cars and other vehicles. The central idea of this project is that the car should learn to drive autonomously from correlating scene information with the actions of the driver. Moreover, we will seek to exploit eye-movement signals to capture driver intention. In the context of this project this system shall be tested and applied in night-vision scenarios with infra-red illumination, which is our main and commercially very relevant application domain. Here we envision a system that can learn to drive a car during daylight and apply the learned control strategies in an autonomous way to the system's augmented field of infra-red night-vision.