Floating point truncation in IsometryMatrix inverse about nalgebra HOT 4 CLOSED

It looks like the printed isometry matrix is exactly the same as your initialization values?
Can you perhaps detail what value you would expect exactly?

from nalgebra.

It looks like the printed isometry matrix is exactly the same as your initialization values? Can you perhaps detail what value you would expect exactly?

So in the first example, the translation values I used as input are 3.7567515671, 6.8521933723, 3.7328615785. But after running the inverse twice I get 3.756751227372396, 6.852192677341882, 3.7328610282357473, here only the first 7 values are the same, ie. 3.756751 which suggest a truncation of f64 -> f32.

For the second example, the same thing is happening but just for the rotation values.

from nalgebra.

This the normal behavior of floating point operations. Inversion, axis_angle, as well as from_axis_angle are all subject to rounding errors due to some floating point multiplication and additions (as well as some trigonometry whenever an angle conversion is involved). So what you are seeing are the rounding errors inherent to floats rather than any f64 -> f32 truncation.

from nalgebra.

This the normal behavior of floating point operations. Inversion, axis_angle, as well as from_axis_angle are all subject to rounding errors due to some floating point multiplication and additions (as well as some trigonometry whenever an angle conversion is involved). So what you are seeing are the rounding errors inherent to floats rather than any f64 -> f32 truncation.

Hmmm, so when I take the matrix from the isometry matrix and use nalgebra Matrix to inverse I get perfect f64 precision. For example:

    let inv_matrix = isometry
        .to_matrix()
        .try_inverse()
        .expect("Pose cannot be inverted.")
        .try_inverse()
        .expect("Pose cannot be inverted.")
        .transpose();
    let rotation_matrix: Matrix3<f64> = Matrix3::new(inv_matrix[0], inv_matrix[1], inv_matrix[2], inv_matrix[4], inv_matrix[5], inv_matrix[6], inv_matrix[8], inv_matrix[9], inv_matrix[10]);
    let rotation = Rotation3::from_matrix_unchecked(rotation_matrix);
    let translation = Translation3::new(inv_matrix[3], inv_matrix[7], inv_matrix[11]);

    let isometry = IsometryMatrix3::from_parts(translation, rotation);
    
    println!("isometry matrix {:?}", isometry.to_matrix());
>>> isometry matrix [[0.8137976526999999, 0.46984630819999984, -0.34202012419999994, 0.0], [-0.44096961619999986, 0.8825640677999996, 0.16317591069999998, 0.0], [0.3785222768999999, 0.0180282984, 0.9254165291999996, 0.0], [3.756751567099999, 6.8521933722999995, 3.7328615784999997, 1.0]]

Here is how to easily reproduce https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=f9f89bde7b9c87a5c27d2f052013e631

from nalgebra.