Are shaping techniques the correct answer the control of visually guided autonomous robots?

In this paper, the authors study how an autonomous robot can learn perception-action (PerAc) associations based on visual information in order to navigate in environments of growing complexity (number of shapes to be analyzed). The starting point is the fact that an association problem (Barto and Sutton, 1981) with a delayed reinforcement signal is NP-complete (Littman, 1994) and so impossible to solve in a general case. However, roboticists have developed a number of tricks to overcome this difficulty. These shaping techniques consist in dividing the learning problem into subproblems which are simpler to learn by the system (Chapman and Kaelbling, 1991).