SMP training

[Xuewang01]

Hi, thank you for the great work on SMP!

I am trying to adapt SMP to a Unitree G1 sim-to-real setting, and I have a question about the prior training setup.

For the prior obs_dict, would you recommend keeping the original observation design in this project, or is it reasonable to customize some parts based on the G1 embodiment, motion data, and deployment requirements?

My main concern is whether modifying the prior observations may improve sim-to-real robustness, or instead break important assumptions in the original SMP design.

Any advice on which observation terms are essential to keep and which ones are safe to modify would be very helpful.

Thanks a lot!

[Murrol]

Thanks for your interest! SMP does not make specific assumptions about the observation design, so it should be fine to modify the prior observations.

Things to be careful about are velocity terms and feature normalization when changing the observation features, since SMP can be sensitive to those changes.