search

isaac-sim / IsaacLab

Unified framework for robot learning built on NVIDIA Isaac Sim
https://isaac-sim.github.io/IsaacLab
Other
1.85k stars 709 forks source link

Weird spawn speed of instanceable usd #669

Open supersglzc opened 1 month ago

supersglzc commented 1 month ago

Hi, I recently converted my robot from urdf to instanceable usd. I double checked if the actual usd is instanceable and printed out the prim.IsInstanceable(). However, I tested the speed in source/standalone/tools/check_instanceable.py and got the following speed:

  Clone time Setup time Memory
convex hull + instanceable + replicate physics 1.5562s 8.4059s 2600MB
convex hull + instanceable 1.6026s 34.429s 2669MB
convex hull + replicate physics 1.3798s 4.2639s 2639MB
instanceable + replicate physics 1.7873s 10.474s 3251MB
instanceable 1.6936s 39.44ss 3339MB
replicate physics 1.4311s 6.7025s 3339MB
-- 1.5059s 227.458s  

Ideally, instanceable should be faster and allocate smaller amount of memory right?

supersglzc commented 1 month ago

@Mayankm96

Mayankm96 commented 1 month ago

Thank you for the investigation. Would be great to provide some more details on the asset itself (number of links, meshes, materials, colliders, joints etc.)

Usually, yes. Instancing should make it faster to parse the scene and consume less memory. However, over time, we have heard from the Omniverse Kit team that they have started optimizing the USD stage implicitly, so some of the benefits of instancing may become less "visible." We tried this out on our end for different assets, and the numbers seem inconclusive (sometimes instancing took less time, sometimes more -- depending on how many meshes, visuals, or colliders are there).

I'll keep this issue open to get some more attention on whether we can just get rid of instancing at some point (if numbers are always promising). Until then, the general recommendation will be to do instancing when you can (and especially if the asset is complex).