Filling the reality gap: Using obstacles to promote robust gaits in evolutionary robotics

In evolutionary robotics, which concerns automatic design of robotic systems using evolutionary algorithms, the well-known reality gap phenomenon occurs when transferring results from simulation to real world robots. Several approaches have been proposed to tackle this challenge, such as improving the simulator, avoiding poorly simulated solutions, or promoting robust controllers by introducing noise in the simulation. In this paper we investigate if the addition of a set of small obstacles in the simulated environment can help promote more robust gaits when transferred to a real world robot. In total 80 robot gaits are tested in the real world, evolved using flat and obstacle-seeded ground planes, and using two different scenario difficulties. The results show that in the baseline scenario the proposed obstacle method has little impact on the reality gap of the evolved gaits, whereas there is a significant reduction for the difficult scenario: The average real world performance ratio is 2.3 times higher than the result obtained with the flat plane, and there are no null-performing gaits.