Out of Time: On the Constrains that Evolution in Hardware Faces When Evolving Modular Robots

With the recent advances of modular robots and low-cost manipulators, the evolution of robots, including morphologies and controllers, has become possible to perform in a physical setup without using any simulators. In this scenario, the evolution cannot be parallelized and the wall time becomes a scarce resource that should be used wisely. This paper analyses different algorithms by using the wall time as a stopping criterion for evolution, and it takes into account that wall time depends on the evaluation time plus the time to assemble and disassemble robots before and after an evaluation. The experiments have been performed in simulation, but the evaluation and assembly time have been carefully modelled from previous hardware experiments. Results suggest that (i) genetic algorithms are severely penalized, (ii) genetic algorithms can be improved by performing several evaluations of controllers for each morphology, and that (iii) evolutionary strategies that can chain several evaluations of robots with close morphologies can outperform other evolutionary algorithms. This finding is not surprising, but to the best of our knowledge previous attempts to evolve modular robots in hardware have not employed evolutionary strategies.