Interior-point solver fails to converge with high friction coefficients

[benthebear93]

Hi @thowell,

I’m currently running some of the older Dojo code to better understand the CI-MPC and dojo-sim papers.

I’ve noticed that the interior-point method often fails to converge when the friction coefficient is relatively high. My guess is that large friction coefficients lead to large decision and slack variables, which makes the KKT system more ill-conditioned and causes the line search to struggle.

I’m wondering if there’s any way to mitigate this. Not being able to use different friction coefficients is a bit limiting in practice.

I know you don’t maintain dojo-sim anymore, but if you have any thoughts or ideas from when you were working on it, I’d really appreciate hearing them.

Thank you!