Challenges for interactivist-constructivist robotics

The interactivist-constructivist (IC) approach to cognitive systems (Bickhard, 1993; Bickhard and Campbell, 1996; Indurkhya, 1992; Christensen and Hooker, 2000) offers a sound framework for understanding cognition and representation and for designing genuinely intelligent artificial systems. Motivated either by theoretical considerations, by the problems of classical artificial intelligence and robotics or by biological inspiration, there is now a substantial body of work that is guided by principles similar to the IC ones, directed towards the development of intelligent embodied agents that learn by interacting with the environment. This body of work, mostly accumulated during the last decade, is probably best represented by the biannual International Conferences on the Simulation of Adaptive Behavior (e.g., Meyer and Wilson, 1991; Nolfi et al., 2006). However, we still lack artificial systems capable of representing the world, in the IC sense, or artificial systems that feature genuine, adaptive intelligence. Why it is so? The purpose of this presentation is to pinpoint the factors that still prevent us to build such systems and to give a speculative suggestion on how we can overcome the present impasse. At least part of the robotics community agrees that the best way to get intelligent robots is not to preprogram them, but to let them learn online, through interaction with the environment, under the direction of a value system. Training may be performed by reinforcement learning, imitation or guidance. A control system that could support this kind of learning should be a collection of parallel, heterogenous, loosely coupled processes, capable of self-organization, such as a neural network (Pfeifer and Scheier,