Self-reflection in evolutionary robotics: resilient adaptation with a minimum of physical exploration

Metacognition is the ability of a system to observe and self regulate its own cognitive processes. In this paper we explore the use of metacognitive processes to improve robot resiliency and learning skills. We examine a robot that contains two controllers: An innate controller that is directly connected to sensors and motors, and a meta controller that monitors and modulates the activity of the innate controller. We show how the meta controller can observe, model and control the innate controller without access to the innate controller's internal state or architecture. Quantitative comparisons of this method with traditional evolutionary robotics techniques show how this form of "self-reflection" is a promising alternative to traditional adaptation methods.