Analytical policy gradient using num diff with contacts

[Jogima-cyber]

Hi,

I'm trying to use MuJoCo to get gradients in contact situations with numerical differentiation (finite differences).
I'm currently using these lines to calculate the forward dynamics :

mujoco.mj_forward(mj_model, mj_data)
mujoco.mj_Euler(mj_model, mj_data)

and then later this line to get the num diff gradients :

mujoco.mjd_transitionFD(mj_model, mj_data, eps, centered, Fx, Fu, None, None)

This works well when there are no contacts, I've checked the gradients.

I was wondering if this is still giving the correct num diff gradient in a contact situation. I've read this issue which seems to indicate that I have to add mujoco.mj_rnePostConstraint(mj_model, mj_data) before the call to mujoco.forward but I'm not sure, what do you think?

[erez-tom]

I expect things to Just Work. Give it a try, and let us know how it goes!