Ideas for reducing RAM?

[kyu-sz]

My training pipeline is tightly bounded by RAM capacity rather than GPU memory capacity. I have a Unitree G1 robot, a table and an object in each environment. I can only run around 512x8 = 4096 environments on an instance which has 1TB RAM. Just wondering if there are any ideas or practices which can reduce RAM usage so that more environments can be launched together.

[alfred-cueva]

You could try reducing the collision meshes through usd modifications, so that collision shapes are minimal

[kyu-sz]

Thank you!

[garylvov]

I can't say I recommend this, but some interesting hacks I found:

#5350

Namely using a large swap will work. I also tried messing around with the memory that suggests some could be unceremoniously freed, but I haven't fully tested it.