Efficiency Through GPU-based Co-Evolution of Control and Pose in Evolutionary Robotics

A key challenge in evolutionary robotics is the computational cost of evolutionary runs. The high computational cost forces researchers to rely on power-hungry computer clusters and, even with these, researchers often are faced with long evaluation cycles that make development of evolutionary experiments a time consuming and tedious effort. In this paper we address this challenge on two fronts. We have developed an evolutionary robotic engine where all individuals are evaluated in parallel using a thread-based implementation on a graphical processing unit (GPU). This engine allows us to run an evolutionary robotics experiment in seconds on a modest laptop. The second avenue of exploration is that we have used this engine to study the role of initial robot poses in fitness evaluation. We find that if we co-evolve initial pose and controller competitively, we can reduce the evaluation period of individuals significantly. Combined the evolutionary robotics engine and the co-evolutionary approach are significant demonstrations of how to make evolutionary robotics more computationally efficient.