I read through all the papers we posted here and downloaded more mostly from IEEE.
None of the papers specify the actual input-output for their NN software, which was odd and moreover it seems the claim that the NN usage does NOT use any of the exact control solutions to quad copters is not valid.
In other words there are exact non-NN analytical (calculus like) solutions for the quad copter UAVs known for sometime. And most of the NN papers for control system of UAVs take advantage of these known solutions.
In and of by itself I have no problem with using parts of the exact solutions for NN version, but the papers are not 100% AI applications without any use of the exact solutions.
This being said the HEX copter will be a totally different solution than Quad, if I follow what is on the papers we posted here! It should have been the same.
Not a problem for us, but it might bruise the ego of the authors quite a bit.
The NN part for control systems is mostly clear now, the code is more or less what was posted by Jean-Louis.
However the input-output and the outer-loop calculation of the motion requires more work. There are two scenarios for this specification:
1. Full training of the NN with little information about the hardware configuration of the copter.
2. Include in the input considerations for the hardware of the copter e.g. drag, arm's length and so on.
I believe the correct solution is #2, while #1 will give an inferior solution.