Evolution of sustained foraging in 3D environments with physics

Abstract: Artificially evolving foraging behavior in simulated legged animals has proved to be a notoriously difficult task. Here, we co-evolve the morphology and controller for virtual organisms in a three-dimensional physically realistic environment to produce goal-directed legged locomotion. We show that following and reaching multiple food sources can evolve de novo, by evaluating each organism on multiple food sources placed on a basic pattern that is gradually randomized across generations. We devised a strategy of evolutionary "staging", where the best organism from a set of evolutionary experiments using a particular fitness function is used to seed a new set, with a fitness function that is progressively altered to better challenge organisms as evolution improves them. We find that an organism's efficiency at reaching the first food source does not predict its ability at finding subsequent ones because foraging efficiency crucially depends on the position of the last food source reached, an effect illustrated by "foraging maps" that capture the organism's controller state, body position, and orientation. Our best evolved foragers are able to reach multiple food sources over 90% of the time on average, a behavior that is key to any biologically realistic simulation where a self-sustaining population has to survive by collecting food sources in three-dimensional, physical environments.