The SEER Project: Robotic Experiments in Subsymbolic Psychology

Chrisley, Ron and Morse, Anthony (2005) The SEER Project: Robotic Experiments in Subsymbolic Psychology. Artificial Intelligence and the Simulation of Behaviour Quarterly (120). 3 and 8. ISSN 0268-4179

Full text not available from this repository.

Abstract

One of the major contributing factors to the success of tomorrows robots will be their ability to learn and usefully adapt to new tasks and situations. Without this fl exibility, robots will remain useful only in highly-constrained and controlled environments. For them to truly act in the real world they must be able to cope with a highly dynamic and changing environment while continually acquiring and refi ning new skills. Such robots must incorporate their experiences into future decision-making processes. Traditional approaches have been restricted by their need for predesigned symbolic representations. This is insuffi cient for the kind of real-world, fl exible, open-ended functionality required of intelligent robots. The COGS approach to robot cognition seeks to avoid this problem. We want our robots to fi nd, for themselves, what the relevant features of a problem are, and indeed what the problems themselves are. More specifi cally, the Seer project seeks alternative forms of representation, ones that will allow a robot to avoid the limitations of having a pre-given, fi xed set of symbolseven grounded ones. In many psychological models, the symbol or concept is the dominant level of description. Unfortunately, for familiar reasons that are too involved to go into here, sole reliance on such a level of description is too restrictive and typically unwieldy in the case of real-world robots. Instead, we use models inspired by data on rat somatosensory cortex, and combine them with established psychological modelling techniques. We then instantiate and test these psychological models of learning in a cortical microcircuit robot controller. We provide our robots with no explicit prior knowledge, and no symbols beyond primitive, grounded sensory inputs and motor outputs. Instead, we give our robots the ability to observe sensorymotor contingencies: invariances in the way actions change the sensory stream. These sensory-motor contingencies are autonomously learned and form the basis of a cognitive model that guides the robots future actions. To allow for more general forms of cognition, we extend this idea of detecting and exploiting sensory-motor invariances to more complex situations.

Item Type: Article
Additional Information: http://www.aisb.org.uk/component/content/article/92-aisb/publications/aisbq/82-aisbqarchive
Schools and Departments: School of Engineering and Informatics > Informatics
Depositing User: Ron Chrisley
Date Deposited: 06 Feb 2012 21:05
Last Modified: 02 Apr 2012 05:09
URI: http://sro.sussex.ac.uk/id/eprint/29461
📧 Request an update