Embodied artificial life at an impasse can evolutionary robotics methods be scaled?

Evolutionary robotics (ER) investigates the application of artificial evolution toward the synthesis of robots capable of performing autonomous behaviors. Over the last 25 years, researchers have reported increasingly complex evolved behaviors, and have compiled a de facto set of benchmark tasks. Perhaps the best known of these is the obstacle avoidance and target homing task performed by differential drive robots. More complex tasks studied in recent ER work include augmented variants of the rodent T-maze and complex foraging tasks. But can proof-of-concept results such as these be extended to evolve complex autonomous behaviors in a general sense? In this topical analysis paper we survey relevant research and make the case that common tasks used to demonstrate the effectiveness of evolutionary robotics are not characteristic of more general cases and in fact do not fully prove the concept that artificial evolution can be used to evolve sophisticated autonomous agent behaviors. Robots capable of performing many of the tasks studied in ER have now been evolved using nearly aggregate binary success/fail fitness functions. However, arguments used to support the necessity of incremental methods for complex tasks are essentially sound. This raises the possibility that the tasks themselves allow for relatively simple solutions, or span a relatively small candidate solution set. This paper presents these arguments in detail and concludes with a discussion of current ER research.