flex-flex collisions

[thowell]

add flex-flex collisions

---

<mujoco model="flex_flex">
  <include file="scene.xml"/>
  <option solver="CG" tolerance="1e-6" timestep="0.001" integrator="Euler"/>
  <size nconmax="40000" njmax="20000"/>

  <worldbody>
    <camera pos="0 -1.5 0.5" euler="70 0 0 "/>

    <!-- Cube 1: Resting on the plane -->
    <flexcomp name="cube1" type="grid" count="8 8 8" spacing="0.07 0.07 0.07" pos="-0.2 0 0.27"
              radius="0.001" dim="3" mass="5.0" dof="trilinear">
      <contact selfcollide="none" internal="false"/>
      <elasticity young="1e4" damping="0.01" poisson="0.1"/>
    </flexcomp>

    <!-- Cube 2: Falling from above onto the corner of Cube 1 -->
    <flexcomp name="cube2" type="grid" count="8 8 8" spacing="0.07 0.07 0.07" pos="0.1 0 1.0"
              radius="0.001" dim="3" mass="5.0" dof="trilinear">
      <contact selfcollide="none" internal="false"/>
      <elasticity young="1e4" damping="0.01" poisson="0.1"/>
    </flexcomp>
  </worldbody>
</mujoco>

mjwarp-viewer mujoco_warp/test_data/flex/flex_flex_cubes.xml --nconmax=8192 --njmax=17000 --engine=warp

flexflexcollision.mov

[thowell]

@quagla the implementation is updated and it looks like the flex-flex collisions are working correctly now.

2 notes:

• there is an issue with resets + warmstarts + flex + the viewer -- not sure what the issue is but can follow up in a future pr
• the existing sap broadphase is silently overflowing

  mujoco_warp/mujoco_warp/_src/collision_core.py

  Line 572 in 130c95b

  if idx < max_pairs:

  as a result we should probably expect some performance regression on the aloha cloth benchmark since we likely need to increase nconmax / naconmax to avoid overflow. can follow up with changes related to that benchmark in a future pr

[thowell]

@kbayes please take a look at the changes to _polytope3. think there is a bug where a vector is not normalized. thanks!

[kbayes]

@thowell oh yes! I had this fixed in an old PR that I closed but forgot to add the fix back in. Thanks!