Thanks for the code! Quick question on the Sim2Real2Sim pipeline

[DanielTakeshi]

Hi @yueyeyuniao thank you for this repository. I enjoyed reading your paper.

I have a quick question about the pipeline. Since you do Sim --> Real --> Sim, is it possible to do this in a repeated cycle? That is:

Sim --> Real --> Sim --> Real --> Sim ...

My main thought is that our overall goal (I believe) is ultimately to improve real world performance of robots, and if we can get a well tuned simulator, then that will mean we can do training in simulation. However, in the current paper, I believe you are only doing one cycle, that is: Sim --> Real --> Sim. But wouldn't we want the improved simulator to then be used to improve the real world robot performance?

[yueyeyuniao]

Hi Daniel, thank you for reading the paper and thinking about the possible extension of the paper.

The Sim2Real2Sim was originally designed to build a high-fidelity simulation environment for flexible object manipulation. After using Sim2Real2Sim, we may do training in simulation and use Sim2Real transfer. So, to improve the real world robot performance, it would be better to use Sim2Real2Sim + Sim2Real or you can say do this in a repeated cycle.