Averaged rototrans bug about pyomeca HOT 7 CLOSED

This is not supposed to work :

def test_rt_mean(rt, seq="xyz"):
    rt_mean = Rototrans.from_averaged_rototrans(rt)

    angles_mean = Angles.from_rototrans(rt_mean, seq).isel(time=0)
    angles_mean_ref = Angles.from_rototrans(rt, seq).mean(dim="time")

    np.testing.assert_array_almost_equal(angles_mean, angles_mean_ref, decimal=2)

If it was that easy to mean a matrix, we would not need to have a proper function.
The reason the first one works, is by chance (probably that using random directly from Euler creates matrices that are near from each other).

The intended behavior is that meaning the angles is not equal to meaning the matrices

from pyomeca.

And the reason they behave differently, which seems surprising at first, is that

angles = Angles(np.random.rand(*size))

does not produce the same as this:

Angles.from_random_data(distribution, size=(3, 1, size[-1]), **kwargs)

which is in the from_random_data from Rototrans

from pyomeca.

This is not supposed to work.

I made the implementation based on this test. What is the difference here?

from pyomeca.

And the reason they behave differently, which seems surprising at first, is that
does not produce the same as this:

Yes, I am aware of that, but in the end it's just random angles and it should work both ways?

from pyomeca.

As I said, meaning Euler angles will get wronger as angles grow.

angles = Angles(np.random.rand(*size))
np.min(angles, axis=2)
np.max(angles, axis=2)

will return values between 0-1 while the random constructor of Rototrans:

angles = Angles.from_random_data(distribution, size=(3, 1, size[-1]))
np.min(angles, axis=2)
np.max(angles, axis=2)

will return values apparently ranging from -20 to 20 (from a quick simulation I made printing the values).

from pyomeca.

Ah ok gotcha @pariterre!
The cumsum in the from_random_data increases the angles too much.

So, as I understand it, our implementation may be good but not the test?

from pyomeca.

That is correct. I don't think there is error in the implementation, but the test won't work as is... A test with know values would do the trick or to use the first implementation (with much smaller angles)

from pyomeca.