english only
EPFL > STI > IMT > LASA > Research > Neural Modeling and Computation > Motion Studies > Normal Imitation
Control and Automation

Neural Modeling & Computation

*Motion studies

  - Normal Imitation
  - Studying apraxia

*Neural Computation

  - Modeling apraxia
  - Modeling movements
  - Demyelination model
  - Frames of reference
  - Visuo-motor imitation

Educational & Therapeutic Devices

Normal imitation

While the neural processes behind vision and motor control are relatively well decoded, the neural processes behind imitation, (i.e. the ability to transfer a visual image into a motor command) are yet poorly understood. Since 1992, brain imaging studies (e.g. Decety et al. Neuroimage 2002, Iacoboni et al. Science 1999, Grafton et al. Exp. Br. Res. 1996) and neurophysiological studies in monkeys (Di Pellegrinoet al. Exp. Br. Res 1992; Rizzolatti et al Cog. Br. Res. 1996) have identified a number of cerebral areas specific to imitation.

Neural Modeling allows to validate neurological hypotheses. Moreover, simulation of the neural processes complement brain imaging studies by providing information on phenomena that cannot yet be explored in brain imaging set-up.

Modelling the neural processes behind human imitation

In this research project, we investigate the neural mechanisms underlying the ability of animals and humans to imitate and to learn by imitation. We develop computational models that give an abstract and high-level representation of the brain areas involved in imitation. The models are grounded in and validated against neurological and behavioral data on human and animal imitation.  

The model gives a high-level, comprehensive, but simplified representation of the visuo-motor pathway behind learning by imitation, from processing real video data to directing a complete dynamic simulation of a humanoid. The model has composite modules whose functionalities were inspired by those of specific brain regions, incorporating abstract models of the Superior Temporal Sulcus (STS), the spinal cord, the primary motor cortex (M1), the dorsal premotor area (PMd), and the cerebellum. Each part is implemented at a connectionist level, where the neuron unit is modeled as a leaky-integrator. The model controls the dynamic simulation through the output of motor neurons, that activate two muscles per degree of freedom per joint. Each muscle is represented as a spring and a damper. Neurons in the PMd module respond to both visual information (from STS) and to corresponding motor commands produced by the cerebellum. The STS-PMd-M1 interconnection is a simplified model of a mirror neuron system.

Numerical simulations

The model is embedded in a three dimensional graphical humanoid simulation of a 37 degrees of freedom (DOF) avatar. Shoulders, hips, wrists, ankles and head have 3 DOFs. Elbows and knees have one.  The trunk is made of three segments with 2 DOFs each. All limbs are attached by hinge joints.  The external force applied to each joint is gravity. Balance is handled by supporting the hips; ground contact is not modeled.

Input data for the imitation are recordings of human motions, see the Motion Studies project. We use 3D motion capture systems, which allow fast and accurate measurement of the movement of ten segments of the body, which allows reconstruction of the whole body motion

Cosimir Simulation
The Cosimir avatars, a 3-D simulation of two humanoid avatars.

Xanim and Vtrainer: A dynamic simulation of a 30 degrees of freedom humanoid developed by  Stefan Schaal

Selected publications

People involved in this project

Last update: 11/03/2008