(Autonomous robots) +(Dynamic environment) +(Intelligent control) =Consciousness?

Owen Holland and Rod Goodman

This talk was based on the observation that the animals suspected of being conscious tend to be those thought to be highly intelligent, and so consciousness may be something to do with intelligence. By analogy with engineering systems, the most 'intelligent' controllers for animals must be adaptive model-based predictive controllers; there may therefore be something in the structure of such model-based controllers that enables a system to be conscious. We outlined a possible strategy for achieving a conscious robot. A simple robot in a simple environment with a simple mission is forced to become more intelligent by making the environment more challenging; intelligence is increased by giving the robot increased powers of building and exploiting models. We proposed methods of spying on the robot's internal models and processes. The idea is to iteratively increase the complexity of the environment, and the robot's ability to deal with it, and to monitor the robot's internal operations with a view to detecting functions and malfunctions characteristic of conscious phenomena. We described a preliminary investigation, in which a simple robot in a simple environment became able to deal with its mission via the building and exploitation of some very simple models. The key stage in enabling consciousness was hypothesised to be the induced splitting of the robot's world model into two linked models, a world model and a self model; we speculated that the self model in such a system will turn out to be a conscious entity.