Question about the role of Domain Randomization and Adapter training

[Leek-LYK]

Hello OpenTrack authors,

Thank you for releasing this great work. I have a question regarding the relationship between Domain Randomization (DR) and the Adapter module.

In the README, you mentioned that it is recommended to enable DR when training both teacher policies and student policies. However, in the paper, it seems that the teacher and student policies are trained without DR, and the Adapter is specifically introduced to handle the dynamics variation caused by DR.

My understanding is that:

• Without DR during base policy training, the Adapter learns to compensate for the dynamics mismatch between the nominal environment and the randomized environments.

• However, if we enable DR during both teacher and student training, the policy itself may already learn robustness to dynamics variations.

Therefore, I would like to ask:

If we follow the README recommendation and train the teacher and student policies with DR enabled, will the Adapter still provide meaningful adaptation ability for dynamics changes? Or does this training setup reduce the necessity/effectiveness of the Adapter?

More specifically:

1. What is the motivation behind recommending DR for teacher/student training in the released code?

2. In your experiments, did you compare:

• Base policy trained without DR + Adapter adaptation
• Base policy trained with DR + Adapter adaptation

3. Which setting provides better sim-to-real performance in your experience?

Thank you very much for your clarification.

[zzk273]

Hello, thank you for your interest in our work.

For the following two settings:

1. Base policy trained without DR + Adapter adaptation
2. Base policy trained with DR + Adapter adaptation

AnyAdapter can work in both cases. However, we have not conducted a dedicated comparison to determine which setting provides better performance.

The reason DR is enabled by default in the released code is mainly to make it easier for other teams to directly evaluate AnyTracker (without AnyAdapter) in sim-to-sim or sim-to-real settings.