Convex Decomposition Approximation makes GPU memory continuously increasing

Dear community,

I wanted to use robots whose collision approximation is set to convex decomposition in OmniIsaacGym. However, I found out that this robot makes the GPU memory continuously increase, and the code will crash. Does anyone know why this happens?

For example, I have run the Allegro Hand task with the following two versions of Allegro Hand, one uses convex hull as the collision approximation method and the other uses convex decomposition.
allegro_hand_convex_decomposition.usd (1.7 MB)
allegro_hand_convex_hull.usd (1.5 MB)

The GPU memory does not increase for the convex hull version but continuously increases for the convex decomposition version, as follows.


I run the code using this script, with headless=True and num_envs=32.

Thanks in advance!

1 Like

Thanks for the report, I did created internal Jira ticket to investigate this, looks like some memory is not correctly freed.
Regards,
Ales

Issue was identified and fixed, thanks for the report. In next release this should be fixed.

Thanks! Is there an expected time for the next release?

The release should happen beginning of May as far as I know.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.