One pain point, not just for ARCore, but for most AR frameworks and languages relates to surface detection and usability in different areas you’re viewing.
I found that the apps work much better at a far distance with a large floor.
Getting it to work on your desk is very hard. Any smooth surface or surface with a blank canvas are difficult to identify and project AR images onto.
Lack of animations
Most of the work is in the 3D models and animation, which needs to be done in specialist software.
The actual programming is a tiny part of it, all built in, and is not easy to create the model and animations in.
It would be simple to add move, spin, scale and other animations manually, which could respond to user interaction but this is not supported in ARCore Sceneform.