Performance differences between v1 and v2 #396
Unanswered
benjamin-macadam
asked this question in
Q&A
Replies: 1 comment 3 replies
-
Hi @benjamin-macadam , In v2 we're using inertias computed from the MJCF loader, but these are pretty unstable with the positional backend without significant tuning. I expect this is likely the issue you're seeing. We expose a parameter to "sphericalize" the inertias here: brax/brax/envs/assets/humanoid.xml Line 20 in 45729a4 which was needed in humanoid. Can you inspect the masses/inertias and see if this fixes your issue? |
Beta Was this translation helpful? Give feedback.
3 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi all,
I'm a big fan of the Brax project and I have been trying to use it for a small project with a masters student, essentially playing around with https://diffmimic.github.io/. We've noticed some pretty big discrepancies between the behavior in v1 and v2 (using the 'positional' backend). We found Brax v1 had pretty reasonable 'rag doll' behavior when you dropped a humanoid into a scene and didn't actuate the joints - in version 2 we've noticed some fairly significant collisions during the initial drop and the character never really comes to rest.
To sum this up as a proper question:
I have found a few very exciting papers in character animation and robotics using Brax at the tail-end of v1. The move to v2 seems like an upgrade, but without any clear guide on how to reproduce v1 behavior in v2 and documentation on how to port code, it will become very difficult for researchers to replicate/build off of these results.
Falling in v1:
2023-09-20.11-25-34.mp4
Falling in v2:
2023-09-20.11-21-48.mp4
To reproduce the script:
Beta Was this translation helpful? Give feedback.
All reactions