Motor Representations for Hand Gesture Recognition and Imitation

We present an approach for grasp recognition and imitation based on models for canonical and mirror neurons, recently found in neurophysiological experiments. Canonical Neurons seem to code object affordances, e.g. possible ways of grasping. Mirror Neurons code goal directed tasks, like precision or power grasping of an object. The major feature of this neuron population is the use of motor information in the recognition step. We propose a Bayesian approach that encompasses all these aspects. Recognition is performed in the motor space and we solve the problem of getting motor information while observing another person. Our approach avoids the complexity of other approaches based on the 3D reconstruction of the hand from images, considering that the hand is a multiarticulated object subject to frequent occlusions. The results obtained illustrate the benefits of designing artificial machines inspired on biological findings and hypotheses, while at the same time, offering robotics technologies as a testbed for such hypotheses.