Reducing GPU memory usage for multi-camera uses

[mike.niemaz]

Here you go:

Orbit multiple cameras (per env) performances optimization

Hello, I’m planning to train skills using the latest orbit(+isaacsim) RL framework. For this I need to add cameras into the scene et then train using those visual tensor data. As a dummy test, I created a simple test code (See below) to validate the fact that I can retrieve visual data as expected. And it works indeed! My concerns are the following: it seems rather slow: around 10 FPS with 2 cameras and 24 environments (num_envs) it gets OOM if I try more than 32 envs on my a RTX A6000 wit…