search

chengxuxin / extreme-parkour

Train your parkour robot in less than 20 hours.
https://extreme-parkour.github.io
Other
472 stars 85 forks source link

Question about the priv_reg_loss update #45

Closed P1terQ closed 1 month ago

P1terQ commented 1 month ago

The code uses history_encoder to predict the priv_latent and the update is in Func:update_dagger. So I have two questions:

  1. Why not update dagger every learn_RL step? Instead, update history_encoder at a frequency of 20 Hz.

  2. Why a temperature coefficient priv_reg_coef is needed? I am also confused about why the history_encoder loss/priv_reg_loss is added to the surrogate_loss?

I can't find the answer to the question in the paper.

Any help will be appreciated!

P1terQ commented 1 month ago

The answer can be in "Deep Whole-Body Control: Learning a Unified Policy for Manipulation and Locomotion "