Environments shared vs separate space and effects on parallelization

[loitonyngo]

In many Isaac Lab projects (e.g., the Leatherback community project), agents are typically distributed across a large space, executing tasks independently (e.g., following unique random waypoints). However, WheeledLab drifting task, all agents seem to train in the same physical space, overlapping with each other.

• I assumed that in the Leatherback example, agents are spread out, likely because their waypoints are randomized, requiring distinct spatial regions.
• In Wheeledlab drifting task, all agents train on the same track (e.g., a fixed circuit), so their positions overlap.
  So I was wondering
• Is this because the environment dynamics (e.g., ground friction, track boundaries) are identical for all agents, making separate spaces unnecessary?
• Or is it a design choice to maximize GPU efficiency by reusing the same physics scene? Has this any effect on the learning rate of agents?

If agents in a task (e.g., a fixed race circuit) must all follow the same trajectory, does distributing them across different spatial locations provide any benefit? Or is it more efficient to train them in a shared space, assuming collisions/interactions are irrelevant?

[geiko246]

Hi, thanks for the discussion! Apologies for the late response. In short, we haven't done any rigorous comparisons between an overlapped environment and separated ones.

There's not a great reason for the overlapping environments aside from the fact that it was a byproduct of how we implemented our spawning and resetting. And actually in the past, at a high number of environments (2000+), Isaac Sim prints a warning saying that too many overlapping collisions may deteriorate the simulation quality. We also haven't added any visual cues to mark the boundaries and so it's perhaps easier to distinguish when it's "painted" by agents.

I think it could be a meaningful improvement to separate the environments while adding in some generated track indicators.

[loitonyngo]

No worries about the "late" answer! In research, you are always busy, and I understand that :)
Thanks for the exhaustive explanation, I see now.
Do you have a general idea of what should be changed in the code to allow you to spawn separate environments?

[geiko246]

I believe I attempted to change it a while ago, but encountered some resistance from the resetting logic. I would start by setting the

WheeledLab/source/wheeledlab_tasks/wheeledlab_tasks/drifting/mushr_drift_env_cfg.py

Line 374 in 501aa2c

env_spacing: float = 0.

to something like 3.0

You'll probably then encounter an issue with the reset manager:

WheeledLab/source/wheeledlab_tasks/wheeledlab_tasks/drifting/mushr_drift_env_cfg.py

Lines 82 to 93 in 501aa2c

	reset_root_state = EventTerm(
	func=reset_root_state_along_track,
	params={
	"track_radius": LINE_RADIUS,
	"track_straight_dist": STRAIGHT,
	"num_points": 20,
	"pos_noise": 0.5,
	"yaw_noise": 1.0,
	"asset_cfg": SceneEntityCfg("robot"),
	},
	mode="reset",
	)

which I think could be fixed by just resetting by referring to mdp.root_pos_w rather than whatever is in there right now.

Happy to help more with this. Just let me know how I can be of further assistance!