False positives can appear

[walseb]

These are two identical convex boxes a long distance from each other and yet collision returns Just True:

(It becomes easier to read with word-wrap on)

λ> collision 99999 ([(1874.9342045000049, (-50.0)),(1874.9342045000049, 50.0),(1974.9342045000049, 50.0),(1974.9342045000049, (-50.0))], polySupport) (([(450.0, (-50.0)),(450.0, 50.0), (550.0, 50.0), (550.0, (-50.0))]), polySupport)
Just True

From my testing it seems as if you have one very high precision (double) convex box and compare that against another one with integer precision (both are using the same datatype), false-positives can appear:

λ> collision 99999 ([(1874.9342045000049, (-50.0)),(1874.9342045000049, 50.0),(1974.9342045000049, 50.0),(1974.9342045000049, (-50.0))], polySupport) (([(450.0, (-50.0)),(450.0, 50.0),(550.0, 50.0),(550.0, (-50.0))]), polySupport)
Just True
λ> collision 99999 ([(1874.9342045000049, (-50.0)),(1874.9342045000049, 50.0),(1974.9342045000049, 50.0),(1974.9342045000049, (-50.0))], polySupport) (([(450.0000000000001, (-50.0000000000001)),(450.0000000000001, 50.0000000000001),(550.0000000000001, 50.0000000000001),(550.0000000000001, (-50.0000000000001))]), polySupport)
Just False
λ> collision 99999 ([(1874.9342045000049, (-50.0)),(1874.9342045000049, 50.0),(1974.9342045000049, 50.0),(1974.9342045000049, (-50.0))], polySupport) (([(450.0000000000000, (-50.0000000000000)),(450.0000000000000, 50.0000000000000),(550.0000000000000, 50.0000000000000),(550.0000000000000, (-50.0000000000000))]), polySupport)
Just True
λ> collision 99999 ([(1874.9342045000049, (-50.0)),(1874.9342045000049, 50.0),(1974.9342045000049, 50.0),(1974.9342045000049, (-50.0))], polySupport) (([(449.9999999999999, (-49.9999999999999)),(449.9999999999999, 49.9999999999999),(549.9999999999999, 49.9999999999999),(549.9999999999999, (-49.9999999999999))]), polySupport)
Just False
λ> collision 99999 ([(1874, (-50.0)),(1874, 50.0),(1974, 50.0),(1974, (-50.0))], polySupport) (([(450, (-50)),(450, 50),(550, 50),(550, (-50))]), polySupport)
Just False

As you can see, simply adding zeroes to the second box won't fix it, but adding something trivial like 0000000000001 seems to. Also if you look at the last prompt, removing the decimals from the first cube makes it work properly again, so it has something to do with the precision

Maybe this is just a inherent problem of the algorithm? I haven't looked into it or tried this with other GJK packages.

[zaidan]

@walseb thx, for finding this bug. The bug was caused by using y = 0.0 for d1 - see: 9c71e25#diff-4f64b18c58764991e6cf79d74e3a8e33R33

[walseb]

Wow thanks for fixing this so quickly!