This is an annoying problem which I also experience a lot. It is bacially caused by the numerical problems in the muscle recruitment solver, and I do have some suggestions to handle it.
I think the root cause is often the pelvis residual forces. In the Ground Reaction Force prediction models, pelvis residuals are implemented as very weak recruited actuators (artificial muscle). They should only be recruited if the model is dynamically out of balance. So the strength of those weak actuators is very low compared to the contract point on the feet.
The problem appears to be the difference in strength between the elements. It causes a numerical problem with the recruitment algorithm. One solution is to increase the strength of the artificial muscles uses by the pelvis residuals:
But this is not a really good solution as it will make the model more likely to use the residuals, even if it shouldn’t.
The other work-around is to change the muscle to recruitment criterion from a third to a second order polynomial. You do this by setting the following in your model:
Main.Studies.InverseDynamicStudy.InverseDynamics.Criterion.Power = 2;
This seems to make the algorithm much more stable, and we are seriously considering changing all the examples to use this as default. It is not likely to influence your results. Some researcher has also suggested that the second order criterion could be better in some cases.
So try that. At least until we can find a way to make the recruitment algorithm more robust to big differences in strength.