The package becomes slower on complicated robots #238
Unanswered
behradkhadem
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Title is self-explanatory but let me elaborate. I'm trying to use this package for reinforcement learning (RL). In RL we use algorithms (lke PPO, TD3, SAC, etc) on the gym environment that has been created. And I'm trying to use this package as environment and I'm glad I've been able to help to facilitate this for everyone. But there is an issue which I'm not sure what to do.
When I'm using the point robot (which has an action space of size 3) I'm able to train with speeds as much as 500 episodes / sec. But when I'm using the panda arm this goes down to 7 episodes / sec max (and it gets reduces to 5 after a while)! And it makes no sense to me. Why this happens? Considering that you need to train in orders like a million episodes to get meaningful results.
At first It became slow due to usage of forward kinematics and complicated computations, but it is not the main culprit. And also, I don't think that increase in degrees of freedom cause this kind of slowness.
One of my guesses is that the usage of meshes in robots like panda causes this. But I have no proof.
What should I try? What can I do?
Beta Was this translation helpful? Give feedback.
All reactions