search

robotology-legacy / mex-wholebodymodel

Matlab MEX interface to the iWholeBodyModel interface.
11 stars 9 forks source link

latest mex-subroutine causes numerical problems #85

Closed Ganimed closed 3 years ago

Ganimed commented 7 years ago

Hi guys,

today I have found by accident a numerical problem that is causing the latest C++ code of the mexWholeBodyModel.

The numerical issue is rising up when the current mex-subroutine is used in combination with a forward dynamics function that is calling the ode15s function of Matlab.

The issue can be reproduced by calling the integrateForwardDynamics.m (matlab/wrappers/+WBM/examples), which calls the fastForwardDynamics() function.

Before I made in the End of January all the changes and optimizations in the C++ code, the integrateForwardDynamics.m worked well without printing any warnings. Now the ode15s method prints a lot of warnings that the values are not meeting the integration tolerances like below.

Warning: Failure at t=8.181408e-01.  Unable to meet integration tolerances without reducing the step size below the smallest value
    allowed (1.776357e-15) at time t.
    > In ode15s (line 730)
      In integrateForwardDynamics (line 55)

So I guess I have something "over-optimized" in the C++ code that is causing this numerical issue. I have to compare the latest code with the previous versions in Git, to localize which change is causing this numerical problem.

If somebody is using the ode15s function and has numerical warnings, it is recommended for the meantime to go back to an older version, until I have found and fixed this numerical issue!

gabrielenava commented 7 years ago

HI @Ganimed, when you fixed this bug in your branch I will test it on WBM_Controller branch to see if everything is ok

Ganimed commented 7 years ago

On the weekend I have found and fixed the bug in modeljacobian.cpp. The corrected version is already uploaded on GitHub. Now the Jacobian works again correctly! You can merge now the correction into the WBM-Controller branch.

But the numerical problem is still not 100-percently solved! If the integrateForwardDynamics.m procedure is now executed, the flood of numerical warnings is not anymore. The situation is now much better, but I get still one or two warnings.

So I was checking the C++ code again. I tried to exchange some parts in the code with a different approach to see if it makes a difference. But nothing has been changed. The changes didn't made any effects. So I went back to an uploaded version state where I was absolutely sure that the Matlab-procedure was working correctly under this "code snapshot". After the compilation I had again the same numerical problem with the ode15s function as with the latest corrected version. So I guessed that the problem is not anymore in the C++ code of the Matlab-subroutine. I have updated everything to the latest version and checked again. But nothing has been changed, the issue was still there. In the end I had the suspicion that yarpWholeBodyModel.h is causing again the troubles. In the header-file I have changed the yarpWholeBodyModelV2 to yarpWholeBodyModelV1 and recompiled everything. Afterwards the numerical warnings were gone!! So my suspicion was right.

Somewhere in a function of the yarpWholeBodyModelV2 interface is causing a numerical problem. In the next days I will try to find out which function in the interface is causing this issue.

To avoid the numerical problems with the ode15s function it is recommended in the meantime to move back to yarpWholeBodyModelV1, until the bug is localized and fixed in yarpWholeBodyModelV2.

traversaro commented 3 years ago

Hi @Ganimed, if this is an issue for you feel free to open it at https://github.com/Ganimed/wbm-library, thanks!