Adaptation and learning in animated creatures

This paper is concerned with synthetic agents interacting with virtual environments, called animated creatures. The animated creatures are articulated graphical figures that are equipped with a set of primitive behavioral patterns. These patterns qualitatively specify which body modules will move concurrently, hence forming a motion group, and which group will move prior to another. The parameterization of these patterns is carried out by the creatures given certain external stimuli. The key to such behavioral adaptation lies in an embedded evolution strategy based selection mechanism. Two examples will be given where this selection mechanism enables a bipedal creature and a six-legged creature to dynamically search for the exact positions as well as duration of body joints as constrained by the qualitatively defined gait patterns. The acquired new stimulus-response pairs are recorded and inserted into a behavioral conditioning network which can be reused and refined during future movements.