Reaching to the Self and to the World: From Babies to Robots

• In Event: The Role of Proprioception in the Development of Reaching and Knowledge About the Self

Abstract

We will present a set of embodied computational models on humanoid robots that address the mechanisms of reaching development in early infancy. Humanoids possess morphologies – physical characteristics as well as sensory and motor apparatus – that are in some respects akin to human bodies and can thus be used to expand the domain of computational modeling by anchoring it to the physical environment and a physical body and allowing for instantiation of complete sensorimotor loops. In particular, we use humanoid robots with pressure-sensitive electronic skins covering large areas of their bodies and thus focus on the somatosensory aspects of reaching development.

First, we develop models of reaching to the self: starting from spontaneous touches to the bodies and corresponding motor-proprioceptive-tactile contingencies. Learning algorithms are used on these data to infer representations of the “body in space”. The learned representations are tested by applying tactile stimuli to the robot bodies and observing the reaching movements.

Second, we study the development of reaching to objects external to the body. These are perceived visually, but we concentrate on the role of haptic feedback in learning the behavior on the robot. If the object is at first randomly contacted, proprioception
provides an alternative to vision to guide subsequent reaching movements. Such “somatosensory coding of space” connects reaching to the body with reaching to external objects.

Third, we will present models of primary proprioceptive and tactile representations on the robot: the robot somatosensory homunculi. The robot performs “motor babbling” or is exposed to tactile stimulations on its whole body. The corresponding proprioceptive (joint angles) and tactile (from the artificial skin) activations are recorded. These are then fed into a self-organizing (or Kohonen) map algorithm and the representations that emerge are studied. Modifications of the standard algorithm that provide the right constraints to channel the learning toward the layout of the neural map observed in primate brains are developed. This work is firmly grounded in collaboration with developmental psychologists and the scenarios the robots are exposed to closely follow concrete findings from behavioral studies in infants.

Author

• Matej Hoffmann, Czech Technical University in Prague
  Presenting Author