Evolutionary Humanoids for Embodied Artificial Intelligence.
The work presented in this thesis aims at investigating the potential of a proposed methodology to create a cognitive control architecture for a humanoid robot. This architecture comprises three hierarchical layers: the reactive layer, the model building layer, and the reasoning layer. The architecture is built on techniques from the field of evolutionary computation, and more specifically evolutionary algorithms. Based on very simple models of organic evolution, these algorithms can be applied to various problems such as combinatorial optimization problems or learning tasks.
The field of artificial intelligence is discussed from a robotics viewpoint. The roles of different paradigms in AI research are considered, and so are the principles of embodiment and situatedness, which are fundamental in the behavior based robotics approach.
Several evolutionary experiments performed on real, physical humanoid robot platforms are presented. These are presented mainly to motivate the use of simulated evolution for control programming of robots. In addition, these experiments constitute a subset of the necessary building blocks of the proposed cognitive humanoid robot architecture, outlined in this thesis. The experiments include sound localization, two instances of machine vision, hand-eye coordination, coordination of actuator motions in a robot foot joint, and two instances regarding learning and adaptivity.
linear genetic programming