Numerical differentiation (or analitycal one) seems to be wrong for the Floating in Contact DAM

I was analyzing a bit the effect of disturbance values for the numerical computation of derivatives. And I found a huge error in the computation of Fx, Lx and Lxx. Unfortunately we don't have unit-test for NumDiff withi this problem. I have attached the code which I used for it test_floating_contact_numdiff.py

The weird thing is that the NumDiff for ABA is working quite well. As I reported here !109 (merged). I also attached this code test_aba_numdiff.py.

changed milestone to %Towards python version 1.0

added bug label

mentioned in merge request !109 (merged)

changed the description

What is this bug and how is it related to !109 (merged)? Does it mean that !109 (merged) cannot be merged? Are you 100% confident that there is no other bugs related to !109 (merged) that are hidden somewhere? Again, I am very reluctant to commit this changes, as the ratio benefit over time-cost and risk is very high. I want a strong commitment that it is working, and I am afraid that, here, "strong" has the direct consequence of implying a lot of validation time.

@nmansard this probably bug is orthogonal (i.e. unrelated) to !109 (merged), you can merge it. That PR fixed some few problems with the unit-test and modify the default disturbance value.

Instead I tried different disturbance values (not using default one) for both DAMFullyActuated and DAMFloatingInContact. I did it in order to understand the numerical error with different disturbance values.

For the DAMFullyActuated case, I encountered the results that I expected (i.e.

$\sqrt{2*eps}\approx 2.1073\cdot 10^{-8}$

). And in fact, I didn't see any rounding error with Numpy or EigenPy.

In order to benchmark better this, I decided to try the same with DAMFloatingInContact. However, I found very big numerical errors in all the disturbance values (including

$``0^{-5}`$

).

I shared both scripts used to test it. I don't discard the possibility of having a bug there. Please try the shared scripts and merger !109 (merged).

I checked and this is not a bug. The problem is in your code. You are using a random state to check the numdiff of floatingincontact. This won't work. You need to account for the shape of the manifold.

I'll close this. Feel free to open it if the issue persists.

closed

@proyan thanks a lot. Could you share the modified code here? I would like to try it and it's nice to share here for future references.

Just use state.rand instead of np.rand to generate the random states.