berkeleyautomation / dart Goto Github PK

Hi,
I have been running the codes of DART, which seems does not agree with the closed solution proposed in the paper. At least, I did not even see the alpha variable as shown in Algorithm 1. Looks forward to your response.
Thanks!

I try my best to understand the paper but I could not find how to inject noise into the supervisor’s policy while demonstrating in a real robot. Could you please tell me when to add noise and how to add noise?
Thanks a lot.

In the "fit" function in "knet.py", you whiten the input X before calling "model.fit". I don't think the "whiten()"function is appropriate for the following reason.
Every time you whiten the input X, the mean and std values are different because they are calculated from different demonstrations. If the size of X is 1, subtracting X by its mean will result in 0 and the std of X will be 0.
Maybe I don't understand your code well enough. Would you please help explain this for me? Thank you.