Real-world reproduction of evolved robot morphologies: Automated categorization and evaluation

Abstract

This paper describes the real-world reproduction of a handful of robots selected from a larger sample of simulated models previously generated by an evolutionary algorithm. The five robots, which are selected by automatic clustering to be representative of different morphological niches present in the sample, are constructed in the real world using off-the-shelf motor components, combined with 3D printed structural parts that were automatically generated based on the simulator models. A lab setup, involving evolution of turning gaits for each robot, is used to automate the experiments. The forward walking speeds of the constructed robots are measured, and compared with the simulated speeds. While some of the robots achieve near-identical results, some show a large performance loss compared to their simulated prototypes, underlining the reality gap issue seen in similar previous works.