Learning by Doing : A Dynamic Architecture for Generating Adaptive Behavioral Sequences

We present a software architecture for the behavioral organization of a mobile robot which is entirely based on nonlinear dynamical systems. The activity of elements from a set of predefined elementary behaviors is controlled over time by coupled differential equations such that complex behavioral sequences are generated. By means of this architecture the behavioral system can be designed such that the robot autonomously plans its actions selecting between multiple behavioral goals specified by an operator. On the basis of predefined local sensor contexts and logical interrelations between the elementary behaviors the system learns to organize these behaviors into a sequence directed towards a behavioral goal. Learning can be run in two modes: (1) the system explores its behavioral space autonomously and (2) learning can be accelerated through guidance by the operator. We prove the feasibility of the approach for sequence generation by an implementation on an anthropomorphic robot and show its planning and learning characteristics by a computer simulation