Inconsistencies in sim2sim transfer to mujoco

[chaitanya1chawla]

I am trying to debug the inconsistencies from genesis to mujoco transfer. I started with a tracking policy, but this had significant divergence, so now I am trying to simply open loop replay a trajectory in both genesis and mujoco.

I mimic all the joint physical properties: kp,kd,armature,frictionloss,stiffness,damping, as well as scene parameters such as dt, substeps etc.
I use affine actuators in mujoco (which should mimic the pd control in genesis).
My xml is also generated from the urdf using mujoco gui.

Am I missing anything? Or is this kind of divergence expected in transfer?

Public Repo for reproduction -> https://github.com/chaitanya1chawla/genesis_sim2sim

Below is the mujoco env, with debug balls showing the states of genesis.

[duburcqa]

Do you enable mujoco compatibility?

            rigid_options=gs.options.RigidOptions(
                [...],
                enable_mujoco_compatibility=True,
            ),

[duburcqa]

Note that Genesis does not advertise matching physics between Mujoco and Genesis. It turns out today Rigid Body physics is mostly consistent at the moment, but there is no guarantee it is still going to be the case in the future.

[duburcqa]

I suggest that you have a look at this unit test, and accompanying gs_sim, mj_sim fixtures.

[chaitanya1chawla]

Hi @duburcqa ,

Thanks a lot for the suggestions!
I went through the unit tests and even extended it to compare all the other physical parameters.

There is still some gap in the recorded joint angles (see picture below). However this gap goes away if I increase the substeps in mujoco.

Any idea about this?

[chaitanya1chawla]

I was earlier using substeps 2 and freq=30, which caused a significant drift.

However, changing substeps to 6 for both genesis and mj fixed this drift.

[duburcqa]

I went through the unit tests and even extended it to compare all the other physical parameters.

I could gladly accept a PR adding these parameters!

There is still some gap in the recorded joint angles (see picture below). However this gap goes away if I increase the substeps in mujoco.

If you can write a unit test relying on gs_sim, mj_sim fixtures that is still failing, I will do make sure we can make it pass successfully.

I was earlier using substeps 2 and freq=30, which caused a significant drift.

If substeps is matching between Genesis and Mujoco, there is no reason to observe a drift. This is weird.

[chaitanya1chawla]

Hi!
Sorry for the delay, got caught up in some deadlines.

I could gladly accept a PR adding these parameters!

I noticed in the newer version of the file, most parameters are already there. But I will do a diff and make a PR if needed.

If you can write a unit test relying on gs_sim, mj_sim fixtures that is still failing, I will do make sure we can make it pass successfully.

Will do this.

In the unit test you mentioned,, for the same xml file, some parameter values are varying between mujoco and genesis, such as links_invweight, links_inertia, dofs_invweight.

I notice the largest gap for the robot's dexterous hands. For example, the fingers have links_invweight ~10^5 in mj and ~10^1 in gs. [although I have added the same joint armature in both - this is verified].

I am wondering if there is a difference between how mujoco computes these compared to genesis, and whether they lead to a significant difference in the dynamics between the two simulators.

Another question, the current behaviour that I am noticing for my policy deployment is that in mujoco the humanoid arms fall a bit in the start of the episode before some torque is generated in them [this is a joint-space trajectory following policy. As the environment initiates the robot in the first step of the reference trajectory, I believe the controller only generates torques on slightly shifting from this reference trajectory]. However this behaviour is not noticed in genesis, and the arms don't fall at all. Both use the euler integrator. I am wondering whether it could be a difference between the implementation of the integrators.

[chaitanya1chawla]

Hi @duburcqa ,
Any idea about the reasons leading to the above differences in parameters?

[duburcqa]

I think it found out what is the issue.

You need to set 'default_armature=0.0' in morph option! I will make it the default when mujoco compatibility is enabled.