Determinism

[erikfrey]

Some of our users would like an option to guarantee determinism (e.g. produce the exact same output every single time it's run with the exact same input) even if it results in a small slowdown.

We will have to investigate how to do this with Warp.

[kevinzakka]

A data point on my end, I cannot obtain deterministic runs with:

wp.rand_init(seed)

[johannespitz]

That would be great to have for debugging.
Do you have a timeline for this feature?

[erwincoumans]

Yes, I'm surprised determinism wasn't a requirement from the start.
Non-deterministic RL is a nightmare to debug, at least a deterministic version that is slower, for debugging purposes would be great.

[thowell]

for now, the requirement for deterministic results is running mjwarp on cpu. this can be done via wp.set_device("cpu"). please see "Is MJWarp on GPU Deterministic?" in the documentation.

[kevinzakka]

Just wanted to bump this up and re-iterate how important simulation determinism is for doing proper science, especially in the context of robot learning!

[mar-yan24]

I would be happy to help contribute to the implementation of determinism if no one has gotten started yet. After doing a brief trace, it seems that support.py has determinism implemented for its matmul. This pattern can probably be extended to the hundred or so wp.atomic_add calls cause the atomics are order-dependent. The next thing after that maybe is to sort contacts/constraints in order after collision is detected? I will probably get started on this in the following week, in which, after an initial implementation, I'll get a PR draft in for atomic changes.