Add convexhull to convehull test

[Levi-Armstrong]

This test triggers the assert for isOutsidePolytopeFace in ComputeVisiblePatch

---

This change is 

[Levi-Armstrong]

This is to assist with issue #326 testing

[Levi-Armstrong]

Here is th ply files if you want to view the convex hull for commit ab76114
link_7_0.ply.txt
sphere_link_0.ply.txt

[hongkai-dai]

Hi Levi,

Thanks a lot for the PR! I think we can fix the problem in that assert with LIBCCD solver. I will file a PR with part of your test to expose the problem.

[Levi-Armstrong]

I have added second of three tests which triggers another assert.

[hongkai-dai]

Hi Levi,

I believe #336 should fix the problem convexhull_convexhull_ccd.

convexhull_convexhull_indep calls a totally different code path. It uses FCL's independent solver with collide function. It incorrectly returns a negative value as the penetration depth. I do not think I can quickly fix that code (I didn't touch collide function or indep solver before). Would you mind to comment out convexhull_convexhull_indep test, and add a TODO in the documentation, saying that we will fix this test later?

I will look into the box_cylinder_test. It triggers !triangle_area_is_zero(a, b, c) assert in faceNormalPointingOutward. This is likely a degenerate case. @SeanCurtis-TRI

Best,
Hongkai

[Levi-Armstrong]

@hongkai-dai I commented out the one test. I also added another test for different scenarios related to checking two convex hulls. Looking at the results of the test it has some large errors when two objects are in collision for distance and nearest points. The following case are currently tested for when the closest feature is:

• Edge to Edge
• Vertex to Vertex

Need to add:

• Edge to Vertex
• Edge to Face
• Vertex to Face
• Face to Face

[sherm1]

/cc @DamrongGuoy

[hongkai-dai]

Hi Levi,

Sorry for my belated reply.

I am looking into your test in https://github.com/Levi-Armstrong/fcl/blob/convexTests/test/test_fcl_convexhull_convexhull.cpp#L542L546. May I ask how you obtain these numbers -0.27552, 0.23776, -0.03776?

In your comment https://github.com/Levi-Armstrong/fcl/blob/convexTests/test/test_fcl_convexhull_convexhull.cpp#L529, you mentioned that the closest feature between the two colliding objects are edge to edge. I think mathematically it is proved that the closest point on the boundary of the Minkowski difference A ⊖ B is an interior point within each facet (see theorem 4.8 of https://github.com/tianxiao/Recent-Advances-in-Real-Time-Collision-and-Proximity-Computations-for-Games-and-Simulations-/blob/master/books/Collision%20Detection%20in%20Interactive%203D%20Environments.pdf). Namely if your polytope is well posed (no two neighbouring faces are parallel), then the closest feature can never be edge to edge, at least one of the feature should be a face.

[hongkai-dai]

Hi Levi,

Sean @SeanCurtis-TRI and I looked into your test on the closest distance. We think in https://github.com/Levi-Armstrong/fcl/blob/convexTests/test/test_fcl_convexhull_convexhull.cpp#L542L546, what you find is not the penetration depth, but the distance between two edges.

The definition of the penetration depth is
min |v|
s.t (A + v) ∩ B = ∅

Namely the penetration depth is the minimal translational distance, such that the two objects A and B can get out of contact. While the distance you find can get the two objects out of contact, it is not the minimal one.

To show the minimal translation distance, Sean draws the two polytopes as

We draw the two closest points p1, p2 found from EPA

And we translate one sphere with a vector p1 - p2, here are the translated spheres. You can see that they are separated

The distance found from EPA is 0.2706222, smaller than the distance you computed (0.27552). So by definition of penetration depth, the EPA distance is more correct than 0.27552.

Note that the closest points found from EPA are not on the edges, and the vector p1 - p2 is not along the x axis.

[sherm1]

Wow! That is an awesome issue comment.

[Levi-Armstrong]

@hongkai-dai and @SeanCurtis-TRI This is great and thank you for the detailed response. I came up with the number using blender and physically measuring two meshes. The one difference is I was measuring along the the same translation vector provided by using primitive shapes assume they should be same. As you pointed out this will not be the same so bad assumption on my part. I will revise the test with the correct information.

[Levi-Armstrong]

@hongkai-dai and @SeanCurtis-TRI This is great and thank you for the detailed response. I came up with the number using blender and physically measuring two meshes. The one difference is I was measuring along the the same translation vector provided by using primitive shapes assume they should be same. As you pointed out this will not be the same so bad assumption on my part. I will keep looking.

[Levi-Armstrong]

So I pulled the meshes into blender and transformed one mesh 0.2m in the x direction. Then created a line between the nearest points(Provided by FCL) to understand where it was on the mesh. Based on your description it sounds like the closest objects should not be two edges but when visualizing it looks like they are both on an edge. Although they could be just very close to the edge but it just looked suspicious. I am going to calculate how close it is and post the results.

Point: -0.025420749082411945, -0.06229177389074456, -0.025420749082411945

Point: -0.032902129264164116, -0.06229177389074456, 0.025420749082411945