The model is biologically inspired in its function, as its composite modules have functionalities similar to those of specific brain regions, and in its structure, as the modules are composed of artificial neural architectures. It is loosely based on neurological findings in primates and incorporates an abstract model of the spinal cord, the primary motor cortex (M1), premotor cortex (PM) and supplementary motor area (SMA), as well as of the temporal cortex (TC) and the parietal lobe (PL) (see figure below). Visual recognition of human movements consists of a retina to egocentric frame of reference transformation in TC [4], followed by a segmentation of the movements in terms of amplitude and speed in PL [7]. Motor control is directed by the spinal cord module and the primary motor cortex (M1) module, both of which have direct connections to motor neurons. Motor neurons activate muscles (2 per degree of freedom), modeled as springs and dampers. Basic motor programs for walking and arm stretching/retracting are implemented as fixed interconnected networks in the spinal cord module (central pattern generators [6]). In the model, premotor cortex neurons respond to both visual recognition of movements and to the corresponding motor commands produced by the SMA. As such, they give an abstract representation of mirror neurons [3,5]. Learning of new combination of movements is done in the PM and SMA modules. These two modules are implemented using the Dynamical Recurrent Associative Memory Architecture (DRAMA) [2] which allows learning of times series and of spatio-temporal invariance in multi-modal inputs.

The model is validated through an implementation in a mechanical simulation of a pair of high degree-of-freedom humanoid avatars reproducing a variety of actions [1]. Data used in the experiments are either video recordings of human movements, or simulated motions of one of the avatars. The experiments showed that the model can learn: 1) combinations of movement involving all joints, including precision movements of the extremities (fingers' movements), 2) complex oscillatory patterns, and 3) sequences with variable timing, as is the case in learning from the human demonstration. The learning algorithm is shown to be robust in the presence of highly noisy data; reliable reproduction can be obtained when the training examples show a high variability in the timing and amplitude of the demonstrated movements.

This project is a part of an ongoing research program that also addresses representations for motor control, motor primitives, and generative perception. See the group's web page http://www-robotics.usc.edu/~agents/imitation.html .

References

[1] Billard, 2000. Learning motor skills by imitation: a biologically inspired robotic model. Cybernetics & Systems, special issue on Imitation in animals and artifacts. C. Nehaniv & K. Dautenhahn editors (In Press). (Compressed postcript version: (38M))

[2] A. Billard & G. Hayes, 1999. DRAMA, a connectionist architecture for control and learning in autonomous robots. Adaptive Behavior , Vol. 7:1. pages 35-64, Jan. 1999. (Pdf version (5M))

[3] G. di Pellegrino, L. Fadiga, L. Fogassi, V. Gallese, and G. Rizzolati. Understanding motor events: a neurophysiological study. Experimental Brain Research, 91:176-180, 1992.

[4] D. I. Perret, M.H. Harries, R. Bevan, S. Thomas, P.J. Benson, A.J. Mistlin, A.J. Chitty, J.K. Hietanen, J.E. Ortega. Frameworks of analysis of the neural representation of animate objects and actions. J. Exp. Biol. , 146, 87-113, 1989.

[5] G. Rizzolati, L. Fadiga, V. Gallese, and L. Fogassi. Premotor cortex and the recognition of motor actions. Cognitive Brain Research, 3:131-141, 1996.

[6] P.S.G Stein and S. Grillner and A.I. Selverston and D.G. Stuart. Neurons, Networks and Motor Behavior, A Bradford book: MIT Press, 1997.

[7] S.P. Wise, D. Boussaoud, P.B. Johnson, and R. Caminiti. Premotor and parietal cortex: corticocortical connectivity and combinatorial computation. Ann. Rev. Neurosci.. 20: 25-42, 1997.