I'll check out Tumbleweed. Any downsides to it compared to Ubuntu forks?
It has been a while, but nVidia drivers have always been a pain to install, especially when you also need an older version of CUDA. If tumbleweed has a better compatibility/easier installation process, it is a big win.
Tumbleweed is rolling release (kinda like arch), although they have a pretty rigorous testing process. So that could be a pro or a con depending on who you’re asking.
If what you’re specifically after is older CUDA toolkit compatibility, then I’d recommend using distrobox instead. That’s what I do for ML workloads. (If you plan on redistributing binaries then you’ll have to strip them with binutils though)
That take depends on what you need from Ubuntu 18.04. I'm not to familiar with how robotics stuff works, but perhaps a docker image would work? That way you can keep whatever libraries you need, and run it on whatever base OS you need. That said, I don't know how much of CUDA or whatever is in the driver vs the userland library, so I'm not sure if it would work.
As for distro, it doesn't matter as long as it's relatively decent. I recommend Linux Mint Debian edition, but I personally use OpenSUSE Tumbleweed.
I saw a question below about Tumbleweed, and you may want to look into OBS, which is OpenSUSE's way of building whatever libraries you need in a repo. So you'd basically find or build a recipe for your version of CUDA and install that alongside whatever else is in the system (assuming the Docker option doesn't work). If you're using a relatively popular stack, chances are someone has already gotten it working.