Xpol
To align these lasers, one method is to shine another 461nm pulse to the chamber and check for fluorescence signal via a photodiode or camera located on the other side of the chamber. Since the beams are generally much larger than the atomic cloud itself, as long as you hit something, it is easy to optimize the signal. For a blue MOT with very high atom number, you can even see a small blue bulb suspending in the mid-air inside the chamber, which could serve as a rough reference to start with.
So now atoms are trapped into the red MOT, you then turn off the red lasers while having the lattice laser (at 813nm) on, so the atoms are loaded into the optical lattice. Using the same method, you take a fluorescence image at the end of the experimental sequence to check if there's any signal. Note that the position of the red MOT depends on both the laser alignment and magnetic fields, so one can either (1) align the lattice beam, in this case, a vertical beam from top to bottom and retroreflected or (2) adjust the magnetic fields to fine-tune the red MOT position. It's an iteration of fine adjustments and looking at images.
Another way to align the lattice to the atoms is instead of shining lattice beams to the atoms, you first send red light through the same fiber. When they hit the atoms in the red MOT, the red light will excite the atoms so you won't see anything now with the fluorescence imaging (as they no longer in the ground state that responds to the blue transition). We call this "the killing beam" as the name suggest. Once you know the rough alignment, switch it back to lattice laser and do the optimizations until you see atoms loaded into the lattice (the distribution and spread of the atom ensemble are different when viewed with camera for atoms that are still in red MOT vs loaded into lattice).
They attributed their major obstacle as their laser phase noise, and they claimed if they subtract the estimated laser noise from the result they would get sub-SQL. But hey, that's not how things work. If you think you can make it, you just make it.
In this paper the lattice is a 1D vertical lattice, so you can think of atoms being trapped in a huge stack of pancakes.
However, there're workarounds to beat this SQL. One approach people are working on is called the "spin squeezing", which, one can think of the uncertainties of a measurement mapped onto a 2D plane, for simplicity, a circle. Spin squeezing is used to improve the uncertainty in one axis while sacrificing the other, like squeezing the circle into an eclipse. There's a group in MIT which attempted to show metrological gain from spin squeezing in 2020, which they failed.