Evolving complex terrain navigation: emergent contour following from a low-resolution sensor

This paper investigates evolutionary approaches to enable robotic agents to learn strategies for energy-efficient navigation through complex terrain, consisting of water and different heights. Agents, equipped with a low-resolution depth sensor, must learn how to navigate between a randomly chosen start/end position in a procedurally generated world, along a path which minimises energy usage. The solution that consistently emerged, was an agent that followed the contours of the map, resulting in near-optimal performance in little evolutionary time. Further, initial experiments with a real robot and Kinect sensor showed that the simulated model successfully predicted the correct movement that would be needed to follow contours. This demonstrated both that the evolved strategies are robust to noise and capable of crossing the reality gap. We suggest that this robustness is due to the use of a low-resolution sensor.