I'm actually doing a bit of that work in my spare time :). Over the years, I've been coming back to this one question: how to make procedurally generated content immersive?
You need procgen to create truly large game universes. But I haven't seen a single game where procedural generation would create an immersive world. Or even an interesting one[0]. In all the games I played, I've seen one or both of the following immersion-destroying things:
- The mechanics are plain obvious. You see a few examples of a generated thing, and it's trivial to guess the pattern and know what you can expect in the future[1].
- There's no story. Stuff is randomly generated, but doesn't make sense. Creatures drop loot they don't use, or shouldn't even have, lore-wise. Dungeons are arranged at random, with no sense of purpose for the layout. Quests are closed loop, with no meaningful impact on anything else in the world.
The closest I've seen to an immersive procgen game is Rimworld, where if you squint just hard enough, you can live the story of your colonists. But if you open your eyes just a bit, mechanics still stick out like a sore thumb.
As an interesting data point, one of the most immersive experiences I ever had with fiction was throwing story prompts at GPT2/GPT3 and going along with the AI narrative. I think it's partly because mechanics of GPT3 are too complex to guess, and purpose/story pieces itself together from correlations in the training dataset. But textual medium also helps - you have to imagine things, and there are no visual inconsistencies to spot.
Right now, I'm casually exploring if the two points above - mechanics and story - couldn't be improved if the models used to generate content were much more complicated in causal sense: instead of few dice rolls, one could build deeper causal graphs and work with probabilities conditional on particular game state. That, and making enough randomizable parameters to ensure the categories aren't obvious.
(If you know of any game, in any genre, that you feel does procgen right, I'd like to know the titles. I'm also on the lookout for related scientific literature.)
--
[0] - Case in point: Kerbal Space Program. Planets and moons are big (1/10th the size of the real things). You can go anywhere and it "feels right". But there's hardly anything to do once you get there. You could, as I tend to, install mods for extended base building and resource mining, but there's only so much fun you can have with it, and the physics engine goes out of its way to make establishing large bases impossible. Celestial bodies in KSPs are very much like checkboxes. After you get there, and tick the "been to $body" checkbox, there's no reason to stay there.
[1] - Example: Starbound. There are random quests, but they follow the same obvious pattern. Find a non sequitur something or someone somewhere, bring it to quest giver, for a non sequitur reward. There are random critters and weapons, but you can quickly tell the categories that are being randomized, and there isn't much variety in those.
As an extreme example, you could complain that the planets in Super Mario Galaxy are way too small, but you'd also be missing the point of games.
Just like adding a hunger/thirst element to Halo doesn't make it more fun.