sebriviere So for example, if estimated focus is position 1247, Fine sequence as is would jump to 1292, then down to 1277, 1262, 1247, possibly confirm 1247 to have best HFD and correct.
The first pass finds the V Curve by using multiple data points (with a least squares fit, the total error is reduced by using multiple points; that is how Gauss helped rediscover the minor planet Ceres, and introducing the mathematical world to the Least Squares method :-).
The second pass looks for the minima by checking the HFD only for a single sample, and therefore has lots of potential error. For the second pass, ASIAIR should just do a slow slew of the EAF (also starting from the right; this will compensate for the backlash) directly to the position computed from the first pass, instead of searching for the local (noisy) minima. Notice that the current ASIAIR second pass usually stops too early (when noise pushes a sample lower than it actually is, or pushes the next sample higher than it actually is).
I have already alerted ZWO of this problem when auto focus first appeared on ASIAIR, and they have confirmed a fix (i.e., by not looking for local minima on the second pass), in addition to changing from a parabola least squares fit to a hyperbola fit (which better matches the optics problem at hand). They had sent me some screenshots of the graphs about ten days ago, and the focusing looked quite reasonable.
Perhaps autofocus could be corrected in the next release.
I had (up to the point the rains came back) myself been using the Average HFD from ASIAIR's Detect Star tool and feeding the HFD to a program that I have written to do a hyperbola fit to discover the true focus location, and then slewing the EAF there myself. I have found the ASIAIR autofocus to be too inaccurate to use, although you find the FaceBook guys loving it, even though their own screenshots show exactly the focusing error that you have described above!
Would backlash be considered for V1.6 too?
EAF Backlash is already taken care of in v1.5, as I noted above.
Viz., notice that ASIAIR always starts way on the right and moving in from right to left. For the second pass it again starts from the right and move to the left. I.e., any backlash is neutralized by this (i.e., they have the same backlashed offsets :-).
There is a further bug that I had (more recently) alerted them to, and that is the EAF scale. With different scalings, the EAF movement can be too shallow (or too coarse) to compute an accurate V curve. When I installed the ZWO EAF on a newly acquired belt driven Askar ACL200, the focus plane moves only by 0.44 microns per EAF step (no kidding). As a result, the V curve from the first pass is completely bogus. The data for the V curve should include HFD values that are above 6, preferably 9 or 10 even.
It is not a universal problem with all belt drives. My WhiteCat51 moves sufficiently (2 microns per EAF step), for example, and it is belt driven. It all depends on the gear ratio and the helical focuser scale and the f-ratio of the telescope/lens. On the other hand, my R&P type focusers all seem to work (both with a ZWO EAF mounted on a FeatherTouch on Borg OTA, and with a ZWO EAF on the Tak FSQ-85's original R&P focuser (4 microns per EAF step)).
So, ZWO also knows of this problem, and we just need to wait for the next fix. In the meantime -- focus like you did before ASIAIR autofocus came out. If you run MacOS, I can send you my program that computes the V curve from the Average HFD numbers from the ASIAIR Detect Star tool, but I recommend just sitting tight and wait for the real fix.