I've been a big fan of REST in the past and it has many advantages over RPC on paper. You can present the client with different Content-Types, Caching, reuse of HTTP status codes, you can guide the client through the API and you can embed documentation in the API if it isn't mostly self-explaining anyway.
But my experience has been that in practice this doesn't hold up and instead you do a lot of unnecessary work to get everything right. Also the HTTP status codes often don't map to your domain logic exactly and using them in your context often feels a bit forced. But the worst thing about REST in my opinion is that you spend a lot of time to design your resources and the interactions they allow. And whenever you do some major additions to your API you hope you find a good solution to add the new functionality and you didn't design yourself into a corner already.
This often feels like a waste of time to me because most of the time I already have a perfectly fine and obvious idea about how to model an API as a set of remote procedure calls. And if I have gone through all this effort to model my problem inside the constraints of REST the next problem is how to call it from the client? Our programs are based on calling procedures so building a good RPC client library is easy, building a good REST client library not so much and in most cases you will just map back from your REST API on the server to a set of procedures in your client library.
Because of this, RPC feels a lot simpler and more natural to me today. What I really miss though is a consistent framework that makes it easy to write RPC services that are self-describing and interoperable. Therefore I created my own project to experiment with new ways to make RPC easier for myself and maybe somebody else finds it useful, too: https://github.com/aheck/reflectrpc