Skip to Main Content
This paper details experiments in which neural network controllers were evolved in simulation that allowed a simple panning camera head to track patterned objects moving against similarly patterned backgrounds in reality. It begins with a discussion of minimal simulations: fast-running, easy-to-build simulations for the evolution of real robot controllers (Jakobi, 1998a; Jakobi, 1998b). The minimeli simulation with which the motion-tracking controllers were evolved (along with the rest of the evolutionary machinery) is then described. Experimental results axe offered in which an evolved controller was downloaded onto a simple panning camera head and satisfactorily tracked two very differently patterned objects as they moved against similarly patterned backgrounds in a random fashion.