Using a reinforcement learning controller to overcome simulator/environment discrepancies

A common approach to simulator/environment discrepancies is to alter simulator designs in order to create a model from which policies are more easily transferable to the real world. We present a different approach which focuses on overcoming discrepancies by designing a controller which is robust to unexpected changes in its environment. This approach is not intended as a replacement for previously developed techniques, but rather as a supplement to them. This combination of discrepancy reduction techniques and discrepancy-robust controllers is shown to be effective in overcoming artificially introduced discrepancies in several simulator-to-simulator transfers, as well as in an actual transfer from a Nomad simulator to a Nomad Scout robot