This is the one-dimensional k-means problem, (only slightly) disguised by
splitting the data into two parallel lines. The solution method is
dynamic programming, which is possible because data can always be grouped
into contiguous sets of points. (The k-means problem in higher dimensions
is known to be NP-hard.)

Reference:

"Optimal k-means Clustering in One Dimension by Dynamic Programming"
 (https://journal.r-project.org/archive/2011-2/RJournal_2011-2_Wang+Song.pdf)

The above paper is referenced in:
"Fast Exact k-Means, k-Medians and Bregman Divergence Clustering in 1D"
 (https://arxiv.org/pdf/1701.07204.pdf)

I used the algorithm from the first paper, which has runtime O(kn^2);
I chose input limits accordingly. The second paper (which I came across
only very recently) claims an O(kn)-time algorithm, which I wish I had
seen first!

My only hesitation about submitting this is that some variation of it
has most likely been used before in a contest. However, the paper
describing the DP approach is in a journal devoted to R programming, so not
one of the mainline CS journals. Moreover, a quick search of Geeks4Geeks
did not turn up a tutorial on this (although there is some discussion of
it in online forums such as stackexchange). But I have an idea for another
variation that I suspect makes it a better choice. Unfortunately I have
not had time to code it. 

ALTERNATE VERSION OF THE PROBLEM:

Assume the customers are all arranged along a CIRCULAR street. I believe
it to still be the case that the minimum sum of square differences can be
achieved by grouping them into sets of contiguous neighbors. (I don't have
a proof of this, but it seems intuitively obvious to me. My intuition has
been known to be wrong.)

Now the "green space" is the interior of the circle and access points can go
anywhere inside.  I don't know how time consuming it would be to do the
pre-computation of the sums of squares, but I think the rest of the algorithm
would just be a modification of the above, with extra rows due to the need for
checking sets that "wrap around" the input data.