There are lots of companies using NixOS for this, BalenaOS (Yocto + Docker), or building their own bespoke tooling on top of a minimal Linux setup.
Although many places start with Ubuntu or Debian in my experience it’s common to invest a lot of time and energy in getting out of that unmanaged setup once the company scales.
The hardware usually comes with vendor-specific libraries (e.g. cuda in the case of nvidia) which are based on a specific version of libc, so then you will have to build your entire alternative OS around that version also.
Which is… never trivial. I’d say 25-50% of my career so far has been repeatedly “fixing” clunky deployments of ROS, OpenCV, L4T, CUDA, cudnn, libc, etc. in Docker and Nix. Fun stuff!
Although many places start with Ubuntu or Debian in my experience it’s common to invest a lot of time and energy in getting out of that unmanaged setup once the company scales.