If you want a sort of FRP-backed datalog, I highly recommend Differential Datalog https://github.com/vmware/differential-datalog It's a datalog flavor/variant that's based on the Timely Dataflow and Differential Dataflow libraries, it's focused on incremental computation over streaming data. Coincidentally it's also used on network switches if that's relevant to your need for deploying on micro-controllers
Where I think this approach fails a bit is the following: There is a language-gap issue where your output-relations need to be mapped in a host language to outputs. And you have to call some functions to collect all data for the input relations. Especially in the context of robotics the latter often is not free of side-effects. Not being able to explicitly control those calls can be problematic.
Having a powerful host-language and explicitly embedding a Datalog-implementation is useful, as the decision procedures for your interactive application are not just a series of if-statements but proper condition-action-rules.
My approach is the other way around. I think embedding strict call-conventions for external functions and explicitly modelling side-effects into the language allows for an expressive LP-language that handles the interactive parts as well as the logic-deductive parts.
But mainly Prolog semantics with side-effects are a whole different beast. This is a different thing I am researching, but basically the whole idea stems from robotics control with Prolog and how handling of side-effects just does not work nicely with SLD-resolution, specifically backtracking over side-effects.
I have a paper on that as well ;)