Hi Jonathan,
thanks a lot for the nice explanation. I’m wondering if you think it’s reasonable to add preconditioning or scaling in acados to improve the numerical stability, since I found that IPOPT (here) and OSQP (part 5 of this paper) both mentioned scaling or preconditioning for better robustness.
As discussed here, I guess sometimes the not-well-scaled problem can also lead to QP solver failure. Although the modeler should try to avoid formulations where some non-zero entries in the gradients are very small or large, having auto scaling might be helpful. Do you think this would be valuable to explore? I want to look into this a bit if you think so.
Best,
Fenglong