Has game AI stagnated?

As indicated by others, this is a pure cost thing. It costs a lot of money and a lot of time to create and tune true complex AI. It costs much less money to tell the artists to create a grotesque but highly detailed monster, rig it for dramatic and stylistic animation, and get the sound guy to put more echo on the RAWR noises it makes but code to effectively behave like "HULK SMASH!" every 20-40 seconds.

The short answer is no. Its just last priority in games. Play FEAR and even today the game holds up. The A.I. will still pull off some amazing things. The Halo series also has good A.I.

In Dark Souls, it would be a lot more fun (less funny, though) if their artificial interesting stopped them from getting stuck on terrain or walking/dodging off of cliffs... Fortunately Dark Souls is designed around exploiting the enemy's very predictable nature and thus the A.I. works as intended; it suits the message system well.

So, er... I guess I agree.

I played F.E.A.R. 1 recently and the A.I. is exactly as you describe. It really elevates the game beyond what it would otherwise be. I think the main problem is most games don't design around having good A.I.

In F.E.A.R., they definitely realised that the good A.I. made the game's primary mechanic - slo-mo - much more interesting, fun, and essential to the game.

Part of this is AI-related, namely having not enough resources dedicated to predicting what will happen to that snakeman if he rolls while on a narrow catwalk, which makes sense because that would be some beastly computation. The big issue you're referring to here, though, has to do with nav meshes, which is information in levels that tells the AI where it's free to move. I think it would be interesting to look at the Dark Souls nav mesh info (which is typically "baked in" to levels and not altered in real-time) and see if the allowances are binary or tiered like "standing is OK here, but please no jumping!"

Not all AI problems are at the coding level. Part of the reason AI doesn't seem to be advancing much is because it takes an incredible amount of animation and voice work to model dynamic AIs who have reactions to large numbers of situations. Take Crusader Kings's AI. It's great. Civilization: pretty good AI. Now go voice, model and animate every interaction that could happen in those games. You're right, no one has the time and money to do that.

So for example in FPS games soldiers are giant, heartless idiots. In real life when you shoot a "bad guy" and they go down, they're probably wounded. That means they will fall and crawl/cry away and one or more of their buddies will probably jump to help them escape the conflict zone. This could be programmed, but it'll take a lot more work to get it into the game, and then you start asking if "it's worth it". Should we give more animations to the AI? or to the player? What makes the game funner? In this example you would need a system that detects injury to specific body parts, animations for each wounded body part, different carry animations for the one helping the victim, and so on and so forth. Lots of work, obviously.

Aside from that there seems to be a rift between what seems like cool AI and what's good AI for a game. Often times I find myself going down a path of "the more human the AI the more interesting for the game" but the problems with that is 1) as humans, we recognize anything that isn't human as just that, and pretty much instantly and 2) much of the fun in many games comes from reliable situations, which humans don't always produce and 3) we spend so much time killing in games I'm not sure I want to be killing the most human enemies possible - an FPS where enemies beg, cry, help each other, run away and die slow painful deaths while crying out to their parents may be an interesting experience, but it's not one we want in most of our games I think.

So other than "AI that moves in real ways and reacts dynamically" and "emotional AI" there's also the problem of "Effective AI", which is both simpler and more complex in my mind. In many games programmers have to "cheat" to make the AI miss - it's pretty easy for a computer to hit a target, as you can imagine, so you could easily have a CoD where the soldiers never miss and always head shot you as soon as that 1 pixel of your pokes over a box - but that ain't fun.

AI is just another part of a game that serves a purpose. Most of the games talked about it's a design issue more than anything. Think about a shooter, you really want every single enemy acting like a human? You against 10-20 humans every couple minutes, you'd never win. Try and separate design from AI. If the AI serves it's purpose, then it's good. It's not about open world jank either, that's typically scripting issues.

Why spend time, money and processing on AI tech when most games are online focused and / or corridor shooters anyway? Especially when it would take away from the visuals.

This is the mentality of some developers.

I think AI in the academic / research sense is different enough from its application in video games that the advances don't translate quick enough.

There is a need and a market for AI middlewear similar to Havok for Physics but it isn't plug and play. I heard for example that Rockstar had to get the euphora devs to come in to help implement that middlewear.