I have not yet push the image comparison test,
For some reason, the test always produce a RMS of 4.77

@image_comparison(baseline_images=['polar_polygon_filling'], extensions=['png'],
                  style='mpl20', remove_text=True)
def test_polar_polygon_filling():
    x = np.linspace(0, np.pi, 5)
    y = np.abs(np.sin(x))
    fig = plt.figure(figsize=(6.4, 4.8))
    ax = fig.add_subplot(1, 1, 1, polar=True)
    # ax.set_rmax(1.5)
    ax.plot(x, y)

Is there any ways to improve the accuracy?

Sorry so we should close the other PR, right?

If the output is correct, then we can close the other PR.
Thanks.

I don't think so; there's nothing wrong with the radar chart in current code. It's only if you apply the other PR that the problem occurs.

My impression is that the fundamental problem is with our API, and none of the related PRs fix that. For non-Cartesian coordinates, one should be able to specify the interpolation policy in a kwarg for each artist, and there should be a consistent and sensible default--presumably the one from rcParams. Then the radar example would simply specify 'interpolation_steps=1, or maybe "connection='straight'" (versus 'curved') for each patch. Or maybe it is a matter of interpolation in "data_space" versus "screen_space".

Cartopy handles this sort of thing, doesn't it?

I changed the title because it was misleading; based on @QuLogic's comment, there was no breakage.

Feel free to re-open if this is still a going concern, but from the above it looks like this is not needed?