Evolving robot vision: increasing performance through shaping

Automated methods for designing robot controllers based on machine-learning techniques have shown great promise when applied to simple robot tasks, but in order to `scale up´ to more complicated problems they will require assistance from human experts, a process that is often called `robot shaping´. In this paper, the difficult problem of learning how to visually track moving objects is examined. It is shown that through the use of shaping techniques, this intractable learning problem can be made solvable. Controllers are evolved in simulation and then transferred to a real robot