Replies: 1 comment 1 reply
-
|
Hi @LyuJZ, Yes, this is a known phenomenon. Please check Section V-A of the paper: https://arxiv.org/abs/2301.04195 TL;DR: SB3 expects the data buffers to be in numpy. This leads to an overhead when converting torch tensors that live on GPU to numpy arrays and vice-versa. |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
From the figure, it seems the sb3 performance is worse than rl_games? |
Beta Was this translation helpful? Give feedback.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi, I am training a policy for object grasping with KUKA arm with IsaacLab. After the same training iterations, I found that the training results of sb3 and rl_games are similar. However, it takes much longer for sb3 than rl_games during the same number of training iterations.
Is this phenomenon normal? If so, do you have any suggestions for improving SB3's training speed?
Beta Was this translation helpful? Give feedback.
All reactions