The problem with this, specifically the assumption that RL gives an equivilance to natural selection and evolution, is that RL typically assumes a computational environment it interacts in while natural selection and evolution assumes the physical world as the environment.
The important difference here is that in order for RL to translate to solving real world problems, you need to faithfully and computafionally simulate the real world's physical processes and rules, or at least enough that n-th order processes exist accurately.
I've done various types of computational modeling and simulation work at different scales throughout my career with all sorts of scientists and engineers and I can tell you, pretty much no domain is there where you have good enough representative models RL can be used in. Some narrow special cases exist but nothing to the degree of a massive environment full of well coupled expert domain models. Some of the best cases are going to be so computationally bound that it would be quicker to do things for real vs simulate.
If you want RL to work and learn, it's likely possible under the connection you point out, but has to do this using physical machines and sensors interacting with the physical world like life as we know it does. Your AGI won't be able to cheat and run through the evolution process quicker using faulty reductionist models we use in most simulations (which is what everyone implicitly is hoping for), IMHO.
If you try this, your AGI is going to learn all sorts of flaws within those environments or at the very least, have so many narrow scoped bounds it won't be that "general." A lot of simulated models are frankly garbage (they have some useful narrow scope but are typically littered with caveats) and they've been in development pretty much since digital computing began.