I also felt the ROS build system more convoluted than necessary; and seemed rather brittle - it was too easy to break it with OS or other updates.

We found that many off-the-shelf ROS nodes didn't do quite what we wanted, and ended up spending much more time than expected rewriting code that we expected we wouldn't need to. It is quicker, and we end up with less & more maintainable code, by writing it ourselves.

I expect this could have been resolved, but when testing ROS we also ended up using more compute resources on-robot than we expected.

Using our own system allows us to build exactly what we require, which has become more important as our system gets larger and more complex; and means integrations into other systems (including testing) are easier.