Thanks

[EelcoHoogendoorn]

Hi Joao,

Thanks for making this available; looks like very clean code, and the multiple shooting LQR makes a lot of sense to me conceptually.

The test cases in the repo are all fairly low dimensional state spaces. Im planning on plugging this into a (modified) brax simulation with len(x)±100. Curious if you expect any trouble there. As long as its a meaningfully differentiable env with appropriately soft contact, it should 'just work' right?

If I get a test working, would you be interested in a pull request?

[joaospinto]

Hi @EelcoHoogendoorn !

Thank you for your interest!

I wouldn't expect a larger problem dimension to pose an issue. I know Lorenzo Amatucci (https://www.iit.it/people-details/-/people/lorenzo-amatucci) has successfully used this for full-body humanoid control on the GPU.

Are you planning on running this on the CPU or GPU? I have some working GPU-friendly LQR implementation in this repo that is not hooked-up by default (I should change the code to just check the device) which would certainly be worth hooking up if you're planning on running this on the GPU.

I do think it should just work; the main risk is probably how well the soft contact logic will work, and while I haven't tried that myself, I have implemented a lot of "soft branching" cost functions for motion planning and things tend to work smoothly.

If there are any missing features that would be useful for you, please also let me know.

Cheers,
Joao

[EelcoHoogendoorn]

Ah that does sound encouraging indeed; I think my problem is a more benign one than a full humanoid.

I noticed the GPU codepath; I will probably get around to using that in the future although I suppose it depends on the amount of batch-parallelism I need. First order of busines would be seeing if its actually a good fit for my problem though.

Itd be nice if someone will implement an IPC contact method like this in JAX/brax in the future; otherwise I might be tempted to do it myself; although im reasonably confident a contact penalty method will be alright for my use case; but its such a nice paper kinda want to play around with it.

[joaospinto]

Sounds great! Will read that paper. Let me know if you run into any issues.