Learning the perceptual conditions of satisfaction of elementary behaviors

Abstract—A core requirement for autonomous robotic agents is that they be able to initiate actions to achieve a particular goal and to recognize the resulting conditions once that goal has been achieved. Moreover, if the agent is to operate autonomously in complex and changing environments, the mappings between in-tended actions and their resulting conditions must be learned, rather than pre-programmed. In the present work, we introduce a method in which such mappings can be learned within the framework of Dynamic Field Theory. We not only show how the learning process can be implemented using dynamic neural fields, but show how the adaptive architecture can operate on real-world inputs while controlling the outgoing motor commands. The proposed method extends a recently proposed neural-dynamic framework for behavioral organization in cognitive robotics. I