tuning solver_ref and solver_imp params to avoid mujoco exceptions

[shivanimall]

While using hollow_cylinder objects I have to fine-tune the solver_ref and solver_imp parameter to avoid the below exception:

mujoco_py.builder.MujocoException: Got MuJoCo Warning: Nan, Inf or huge value in QACC at DOF 9. The simulation is unstable.

This happens when a random action is applied to the object or the object accidentally falls down. I am facing a lot of these exceptions which is impacting my downstream tasks.

Do you have any suggestion/documentation/algorithmic approach on how to tune these based on objects' attributes?
Applying brute-force fine-tuning approach is quite intensive in my case.

Many thanks, please let me know if you need more details.

[yuvaltassa]

Hi,
Sorry for the delay. MuJoCo's current defaults are quite soft, which can lead to a lot of penetration and subsequent large forces and divergence. I t will probably help to stiffen everything globally by reducing solref[0] for all geoms. Another possibility is to reduce the time-step.

<default>
  <geom solref=".004 1"/>
</default>

[shivanimall]

Hi @yuvaltassa thanks for the response. I tried reducing solref for all geoms to [0.004, 1.], plus trying control time-step range from 1/20 to 1/100 but facing the same issue. To note, I have composite objects with different sizes in different episode runs. (I create procedurally generated objects in robosuite, as here)

Would I need to further fine-tune this param based on object size? Many thanks for your help! (also, posted as follow-up on #98 )

[yuvaltassa]

These timesteps are still quite large. Did you try anything in the vicinity of 1ms?

Another thing to try is to switch the integrator from Euler to RK4.

If neither of these work for you, please post an example model where the failure is reproducible and we'll continue to investigate.