Tackling simulation inconsistencies in the robot design process by selective empirical evaluation.

We present a computational design pipeline that allows evaluation of the robot and environment parameters in a robust manner, giving insight into interactions that can lead to mismatch between simulated behaviour and reality. Our pipeline evaluates robot designs across different design parameters in a large variety of stochastically-defined environments to robustly infer the qualitative effect of robot parameters on its performance. We then quantitatively ground this insight by selecting and building a small number of physical robots to help establish bounds on the trend in parameters observed in simulation. This combination of simulation and empirical evaluation helps narrow the sim-to-real gap without excessive expensive physical testing to augment the intuition of the human designer.