Bug in RealityKit Transform

Question

kuprel OP

Created Jul ’21

Replies 2

Boosts 0

Participants 2

The same rotation produces different rotation matrices. For example this:

let R1 = Transform(rotation: simd_quatf(from: [1, 0, 0], to: [0, 1, 0]))
let R2 = Transform(rotation: simd_quatf(from: [2, 0, 0], to: [0, 2, 0]))
print(R1.matrix)
print(R2.matrix)

outputs

simd_float4x4([[-1.1920929e-07, 1.0000001, 0.0, 0.0], [-1.0000001, -1.1920929e-07, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 1.0]])
simd_float4x4([[-3.0000005, 4.0000005, 0.0, 0.0], [-4.0000005, -3.0000005, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 1.0]])

With some cleaning up the matrices look like this

// R1
[0, -1, 0, 0]
[1,  0, 0, 0] 
[0,  0, 1, 0]
[0,  0, 0, 1]

// R2
[-3, -4, 0, 0]
[ 4, -3, 0, 0] 
[ 0,  0, 1, 0]
[ 0,  0, 0, 1]

We see that R2 actually scales its input. A pure rotation should have a scale of 1.

Another disagreement occurs when evaluating the Transform's scale. For example this

print(R2.scale)
print(Transform(matrix: R2).scale)

outputs this

SIMD3<Float>(1.0, 1.0, 1.0)
SIMD3<Float>(5.000001, 5.000001, 1.0)

Why are the scales different for the same exact transform?

Also, the rotation angles do not agree. This

print(R2.rotation.angle * 180 / .pi)
print(Transform(matrix: R2.matrix).rotation.angle * 180 / .pi)

produces

90.00001
126.8699

90 degrees is the correct rotation angle. Where does 126.87 degrees come from?

Boost

Answer 1

Graphics and Games Engineer OP

Apple

Jul ’21

The Transform API specifies that rotation should be a unit quaternion. The quaternion you used for R2 is not normalized.

0

Answer 2

kuprel OP

Jul ’21

I see that it says that. But the fact still remains that it is possible to create a Transform that disagrees with itself.

This test should never fail:

func testTransform(A: Transform) {
    assert(A.rotation.angle == Transform(matrix: A.matrix).rotation.angle)
}

Something strange is going on internally with RealityKit's Transform. There must be redundancy in Transform's state that leads it to disagree with itself when passed an unnormalized quaternion.

0