Making MPM simulation more realistic for humanoid locomotion

[jnskkmhr]

Hi, Newton team.
I have been working on humanoid locomotion on deformable terrain and I want to use newton as a part of simulation demonstration.

So far, I managed to run 29dof Unitree G1's RL based locomotion policy on MPM sand.
However, sand is not deformed that much and the foot is not getting large enough resistance as opposed to real-world experiment using other bipedal robot.

Is there any way to improve simulation reality?
One example I can think of is to use a visual triangular mesh as a rigid body collider in MPM solver, but I do not know how to change MPM example code to do this.

newton/newton/_src/solvers/implicit_mpm/solver_implicit_mpm.py

Line 1311 in a4eaa35

collider_meshes: Warp triangular meshes used as colliders.

I think ANymal example uses fairly accurate collider, so I am wondering if using visual mesh as collider improve simulation reality.

g1_mpm_locomotion_isaac.mp4

hector_gm.mp4

[jnskkmhr]

Here is code to run G1 simulation shown above.
https://github.com/jnskkmhr/newton/blob/main/newton/examples/mpm/example_mpm_robot.py