How does virtual training translate to physical model performance?

Hello everybody, I am currently participating in an AWS sponsored DeepRacer challenge for my education program. Each of ten teams gets 100hrs or AWS sponsored server time. A month from now each team will submit one model. The models will compete on a physical track, set up and managed by AWS people. The track will be re:invent2018. I have dug deep into the internet's ressources regarding DeepRacing and I have gathered lots of information. But 99% of that information is for virtual races. Can you please direct me towards sources on physical DeepRacing? Specific questions are: General things to consider when translating from a virtual to physical environment How will speed translate? I have heard that physical racers will have a speed dial to set it's speed. How will this relate to the speeds the model was trained on? I have heard that physical racers will not support asymmetric action spaces. Will there be issues with a model that was trained with an asymmetric action space? Thanks for any tipps and links, Max :)