when a Shape's center is inside a MeshShape, they're not reported as colliding

[KelseyHigham]

demonstration

mesh-capsule-collision-bug.mp4

code for that demonstration

--                             ▄                       █      
--    ▄▀█ ▄▀▄ █▀▄ ▄▀▄ █▄▀ ▄▀█ ▀█▀ ▄▀▄    █▀▄▀▄ ▄▀▄ ▄▀▀ █▀▄    
--    ▀▄█ ▀█▄ █ █ ▀█▄ █   ▀▄█  ▀▄ ▀█▄    █ █ █ ▀█▄ ▄█▀ █ █    
--    ▄▄▀                                                     



local vertices = {
  {0, 0, 0},
  {1, 0, 0},
  {0, 1, 0},
  {0, 0, 1}
}
local indices = {
  1,3,2,
  1,2,4,
  1,4,3,
  2,3,4
}
local vertex_format = {
  {type = 'vec3', location = 'VertexPosition'}
}
local index_format = {
  {type = 'index16', location = 'VertexIndex'}
}
local triangle_buffer = lovr.graphics.newBuffer(vertices, vertex_format)
local index_buffer = lovr.graphics.newBuffer(indices, index_format)
local world = lovr.physics.newWorld()
local tetrahedron = world:newMeshCollider(vertices, indices)
tetrahedron:setPosition(0, 1, -1)
tetrahedron:setGravityIgnored(true)

capsule = world:newCapsuleCollider(0,0,0, .05, 1)
capsule:getShapes()[1]:setSensor(true)

left  = world:newSphereCollider(0,0,0, .05)
right = world:newSphereCollider(0,0,0, .05)



--      █               ▄▀▄     █ █ ▀   █            
--    ▄▀█ █▄▀ ▄▀█ █ █ █ █   ▄▀▄ █ █ █ ▄▀█ ▄▀▄ █▄▀    
--    ▀▄█ █   ▀▄█  █ █  ▀▄▀ ▀▄▀ █ █ █ ▀▄█ ▀█▄ █      
--    from Josip



function drawCollider(collider, pass)
  local shape = collider:getShapes()[1]
  local shapeType = shape:getType()
  local x,y,z, angle, ax,ay,az = collider:getPose()
  -- draw shape at collider's position with correct dimensions
  if shapeType == 'box' then
    local sx, sy, sz = shape:getDimensions()
    pass:box(x,y,z, sx,sy,sz, angle, ax,ay,az)
  elseif shapeType == 'sphere' then
    pass:sphere(x,y,z, shape:getRadius())
  elseif shapeType == 'cylinder' then
    local l, r = shape:getLength(), shape:getRadius()
    local x,y,z, angle, ax,ay,az = collider:getPose()
    pass:cylinder(x,y,z, r, l, angle, ax,ay,az)
  elseif shapeType == 'capsule' then
    local l, r = shape:getLength(), shape:getRadius()
    local x,y,z, angle, ax,ay,az = collider:getPose()
    local m = mat4(x,y,z, 1,1,1, angle, ax,ay,az)
    pass:cylinder(x,y,z, r, l, angle, ax,ay,az, false)
    pass:sphere(vec3(m:mul(0, 0,  l/2)), r)
    pass:sphere(vec3(m:mul(0, 0, -l/2)), r)
  end
end



--              █      ▄         
--    █ █ █▀▄ ▄▀█ ▄▀█ ▀█▀ ▄▀▄    
--    ▀▄█ █▄▀ ▀▄█ ▀▄█  ▀▄ ▀█▄    
--        █                      



function lovr.update(dt)
  -- position capsule between hands
  local leftPos  = vec3(lovr.headset.getPosition('left'))
  local rightPos = vec3(lovr.headset.getPosition('right'))
  local delta = rightPos - leftPos
  if delta:length() > 0 then
    capsule:getShapes()[1]:setLength(delta:length())
  end
  capsule:setPosition((leftPos + delta:mul(0.5)):unpack())
  local orientation = quat(vec3(delta):normalize())
  capsule:setOrientation(orientation)

  left:setPosition(vec3(lovr.headset.getPosition('left')))
  right:setPosition(lovr.headset.getPosition('right'))
end



--      █                  
--    ▄▀█ █▄▀ ▄▀█ █ █ █    
--    ▀▄█ █   ▀▄█  █ █     



function lovr.draw( pass )
  pass:setCullMode('back')
  pass:mesh(triangle_buffer, index_buffer, mat4(tetrahedron:getPose()))

  pass:setDepthWrite(false)

  pass:setColor(0,1,0,.5)
  if world:collide(tetrahedron:getShapes()[1], capsule:getShapes()[1]) then
    pass:setColor(1,0,0,.5)
  end
  drawCollider(capsule, pass)

  pass:setColor(0,1,0,.5)
  if world:collide(tetrahedron:getShapes()[1], left:getShapes()[1]) then
    pass:setColor(1,0,0,.5)
  end
  drawCollider(left, pass)

  pass:setColor(0,1,0,.5)
  if world:collide(tetrahedron:getShapes()[1], right:getShapes()[1]) then
    pass:setColor(1,0,0,.5)
  end
  drawCollider(right, pass)
end

[bjornbytes]

After some investigation in chat this seems to be an intentional behavior of ODE. ODE doesn't support the case where the center of mass of an object gets inside a triangle mesh. It's sort of like triangle meshes are considered "lists of triangles" instead of a solid object? The behavior is still surprising and not very friendly, to a point where this should be considered a bug. Some thoughts:

• If no action is taken, this discovery should at the very least be documented as a limitation of the physics module.
• ODE could be modified to detect intersections with triangles even when the shape's center is behind the triangle.
  • I think this would still fail to detect if a shape is entirely "inside" a mesh shape, so it's not really sufficient.
• ODE has another shape which is a convex hull, which at a glance appears to be more efficient/robust.
  • If this replaced MeshShape then there would be a regression in functionality.
  • It could be added as a new Shape instead.
• An attempt to "glue" together 2 mesh shapes with different windings was unsuccessful (ODE complained about mass)

[KelseyHigham]

i think doing a combination of 1 and 2 would be a big improvement. i'm curious why ODE doesn't do 2 by default – would it mess up the normal calculation, and push shapes towards the inside of the mesh?

[bjornbytes]

JoltPhysics has a mBackFaceMode option that can be set when doing collision queries. It can be IgnoreBackFaces (the default) or CollideWithBackFaces 👀

[bjornbytes]

LÖVR uses Jolt now. MeshShapes still aren't treated as solid objects, by design. However A) there is a ConvexShape which is a solid shape that can be created from a triangle mesh and B) raycasts/shapecasts treat MeshShape triangles as being double-sided instead of single-sided.