Increasing the number of environment columns/rows improves performance?

[vassil-atn]

Hi,

As a note, I'm using the legged_gym implementation for training a quadruped but for what it's worth I believe it should also be reproducible in the anymal_terrain.py implementation here.

I am trying to create a terrain consisting of a flat plane with 5 boxes of different height on it. In that case I need terrain.env_rows (num_rows) to be 5. In theory terrain.env_cols should be 1, as I only have one type of terrain, right? However, I've noticed that increasing the number of columns takes up more memory (as expected), but highly improves the collection time when training my algorithm. I'm using 4096 agents.

It seems to me that having multiple robots start at the same point reduced the performance of the simulation. Is this really the case and if so, what could be the reason for it?