Self-organization of behavioral primitives as multiple attractor dynamics: a robot experiment

Abstract

I investigated how behavior primitives are self-organized in my previously (Tani, 2001) proposed "forwarding forward model" neural network model in the context of robot imitation learning. The model is characterized with the so-called parametric biases which adaptively modulate for embedding different behavior patterns in a single recurrent neural net in a distributed way. My experiments, using a real robot, showed that a set of end-point and oscillatory behavior patterns are learned as fixed points and limit cycle dynamics respectively with adapting parametric bias for each. Further analysis showed that diverse behavior patterns other than learned patterns were also generated because of self-organization of the nonlinear map between the parametric biases and behavior patterns. It is concluded that such diversity emerges because primitives are represented distributedly in the network.