Periodic nonlinear principal component neural networks for humanoid motion segmentation, generalization, and generation

Abstract

In an experiment with a soccer playing robot, periodic temporally-constrained nonlinear principal component neural networks (NLPCNNs) are shown to characterize humanoid motion effectively by exploiting fundamental sensorimotor relationships. Each network learns a periodic or transitional trajectory in a phase space of possible actions, and thus abstracts a kind of protosymbol. NLPCNNs can play a key role in a system that learns to imitate people, enabling a robot to recognize the behavior of others because it has grounded that behavior in terms of its own bodily movements.