State-of-the-art robotic technologies rely on preprogrammed models and largely
lack the capacity to adapt to unexpected changes to their bodies or the
environment. Furthermore, they often blindly execute their movements and their
perception of contacts with the environment is absent or
restricted to the end-effector. Humans, on the other hand, seamlessly control
their complex bodies, adapt to growth or failures, and use tools.
Utilizing redundant multimodal sensory information plays a key part in these
processes. However, the mechanisms of how the brain represents
and updates the properties and current state of the body (so-called body schema
or body image) are still not fully understood. Therefore, modeling and
implementing body representations in robots can serve to crystallize known
biological facts and simultaneously design robot controllers which can
perform with the reliability and adaptability of humans. In this talk, we will
briefly review the concept of body schema in biology and robotics.
Then, we will present case studies on learning body models in robots. First, we
report a quantitative analysis of sensorimotor flows in a
running quadruped robot using tools from information theory and show how the
agent can use this to acquire a primitive body schema. Second, we present
case studies in the iCub humanoid robot that capitalize on whole-body artificial
skin to perform self-calibration and to develop a safety margin
around its body through visuo-tactile associations.