What’s the deal with Modern Racing Games and their AI?

It wasn’t all that long ago where we ran an article displaying a collection of videos from almost all modern racing sims, which despite running on completely different engines created by five very different developers, displayed roughly the same offline AI issues.

We’ve got more to add to that list. It’s definitely a hard pill to swallow when you realize that racing games have regressed in quality over the years. Each of these games was promised to be something special, complete with their own group of diehard fans patiently anticipating the inevitable release date, and instead the end result was a dismal display of technical ineptitude:

The Golden Age of Sim Racing? Nah.


34 thoughts on “What’s the deal with Modern Racing Games and their AI?

  1. That rallycross bug was fixed a long time ago, and the game was (and still is) in Early Access, so it’s hardly surprising there are bugs, I’d focus on the released games.


    1. Eh, if something I never use is broken, it’s not gonna influence my decision much. I’m used to games including content they shouldn’t’ve cause they could only half ass it, and the rest of the game works fine.


  2. For the Forza 6 one; it’s almost certainly a pathing issue. As far as I know/I’ve seen there isn’t any implemented difference for the AI pathing between a wet and dry track, so AI will take the given path; and be damned if there’s a large puddle right on line that’ll make the cars aquaplane off.

    Turn 10 have actually acknowledged there are AI issues (more so what I call the “Mercedes AMG F1 Effect”), have stated that it’s been tweaked it a little, and are “investigating further”.


    1. It’s also an issue that’s specific to that track, high-powered cars, and rain. Everywhere else the AI is fine, although in Single Player the lead AI is much faster than the others.


  3. Money talks and Dev want money, that’s why they release “Beta” state and all that crap.
    Wish i started simracing in 2005 and not 2009


    1. Or maybe because it still takes a lot of time to build a sim racing game, and if there’s just an indie dev company behind it, then will take longer. When you say they want money, I see it more as a necessity to keep their company and staff going while developing the game. Sim racing studios don’t start with 20M. They need to release earlier and get money along the way.


  4. All the sim’s AI is equally shit? LOL hardly.

    Assetto Corsa still takes the cake for having the most terrible AI seen in a racing sim in the last decade. Kunos deserves all the shit they can get for that hopeless AI code they’re fumbling around with.


  5. This is really getting pathological. If you have nothing better to do than all day long picking weird AI situations (whereas in general this works just fine, be it AC, PCars, rfactor 2… I can tell because I actually play races with AI quite often… yes there are flaws sometimes, as in EVERY F*** game, furthermore if complex physical interactions are modelled which is specifically the case in simulators), that’s just sad. If you can’t deal with it: Get professional help! Find a proper mental institution! I mean it.


    1. One of forums I took part used to run an offline (AI-based) series using LFS’ Formula car. In the second season, it was so hectic that LFS’ AI idiocy caused mass penalties, DSQs, and injuries, to the point a driver have to be killed off due to a huge incident involving him and another racer.

      That series switched to F1C and rebooted after that.


      1. The AI needs to be “trained” in each car and on each track by running laps. While this doesn’t at all facilitate pick up and race gameplay, once the AI knows the track, it is some of the most believably dynamic competition I’ve seen in a single player mode.

        I’m not sure if it was done this way to allow the AI to grow with the player or if there is something inherent with the “learning” process but if someone could find a way to skip that step, I imagine it would change how people play offline.


  6. I think the key to understanding why we’re getting these issues is knowing how the damned AI works in these games. Last time I delved into understanding how an AI drives was in NFS3, where the track had a few “rails” the AI cars followed. I think even the physics were different for those cars (surely when pushing against you they would have tons of grip). We ARE however talking about an arcade racer for the PC with a 2-point physics model (front and rear, not even computing all four wheels), not a full-blown sim.

    Issues were present even in Race07 and GTR2. In Race07 rain started one lap before the finish. The AI threw all the cars into the pits to change tyres. Silly buggers they were. I’ve seen another glitch in R07 where one AI would drive on three wheels (the fourth had come off). I’ve only experienced these ONE time, and they were funny rather than frustrating, because 99% of the time they would work correctly. They weren’t perfect, but they worked well for so much of the time that these freak accidents were just that — funny glithes that added to the fun.

    In GTR2, in a night race, the AI would drive with both headlights broken, which is dangerous and illegal. It was a bit unfair to the player and not entirely in the endurance racing zone, but for such a complete and well-done game it was something I could ignore easily.

    You could argue that Kunos has little experience with AI, but other producers are having issues as well. R3E is using (as I hear) an older engine, SMS should have experience from GTR2 and the Shift games and Codemasters… what the hell happened to Codemasters?

    Could any of the readers here that have some modding and/or development experience shed some light on how these virtual bastards drive their cars? Is it just sloppy programming or is it harder than before to teach a computer to drive a pretend race car? I’m not defending anyone. I’m an R3E fan, yes, but not a fanboy. I’m just curious as to how they managed to all fuck up in the same way with so many different games.


    1. I don’t have experience writing AI for racing sims in particular, but in principle there are some main points:

      In a simple game with 1 point of contact with the ground and the only difference between cars is their acceleration you just need to know how sharp a corner is and you can calculate how fast the car needs to be going to drive around it safely. If there’s an obstacle in the way, it changes how sharp the corner is and the AI changes speed.
      In a complicated game where every car has maybe 500 physics parameters, no 2 cars corner the same way.
      The same applies to tracks – when every corner is not just a different shape, but cambered, sloped, bumped, etc. with a unique physics mesh, you can’t assume a certain amount of cornering force applies anywhere with the same radius.
      This means that instead of a linear amount of work to make the content work (test every car once, test every track once, to be sure the line the AI chooses is correct) you have quadratic (every car * every track) or worse (every car * every track * every weather scenario), and when you add in car vs. car scenarios it’s plain impossible to test every possibility – what if a car’s stopped on the track? Well, ignore that because that’s conditions for a full course yellow so the AI doesn’t need to be fast. What if someone is cornering 40% slower than the AI? It’s questionable whether that’s racing, but the AI should navigate around them.
      Since players vastly outnumber developers, it’s likely they’ll find a situation that the developers have never seen. A dev just doesn’t have time to watch laps at every course, in every car, after every patch. Whereas 10,000 players will pick all sorts of things.

      It’s relatively easy to teach a virtual dude to drive a car around the course. Driving at the limit takes actual skill (hence why most sim racers cannot do it). If the dev team has that skill, they’ll be able to identify where the AI has problems, but it does require a higher level of smartness in the AI code to attain it consistently.
      This is the sort of thing all sorts of robots will struggle with – they can complete the task, but can they predict the most efficient option? Much harder.
      Most games just uncap the limits of the car, and let the AI have extra horsepower if it pushes the throttle to 110%, extra grip when it turns the wheel harder, etc. This gives them reasonable laptimes without having to drive as well as humans can.

      Human drivers know which corners are tricky before they’ve even played the game. There’s no mechanism for transfer of AI knowledge from one game to another – unless the developer is reusing their AI code, the AI knows nothing about driving, like a player who’s never been in a driving game before.
      You can split this into learning AI, and non-learning AI. With a learning AI, after every lap, it tweaks the parameters that led to it deciding how to drive, and compares the results. The problem here is overspecialization and the length of the development cycle. To get a seriously fast laptime ‘fairly’ you need an intimate knowledge of the circuit – exactly what throttle & gear & brakes works in each corner at each point on the track. This takes time to learn – which means, if you change a car’s physics, you need to teach it to drive everywhere again. If you change the game’s physics (NTM v6 or whatever) then you need to teach every car to drive again. It’s also hard (as in, computer scientists would love to find a way) to generalize the knowledge the AI has so that it can learn other tracks quicker in the same car. Right now the best option is to tell it what good driving looks like, so that it knows which corners it’s driving well and which ones it can speed up, which at least gives it a direction to try changing.
      Non-learning AI is just told exactly how to respond to any situation. This is how most games work, because it’s that much easier to actually finish the product. If you’re in situation X, respond this way. Your tires are sliding? Let off the brake pedal until they are rolling again. Not turning in well enough? Tap the brakes or lift the throttle for a little bit. The AI doesn’t have to know that it’s doing weight transfer, the programmed response is chosen to do that.


  7. It is no real surprise they are all in a similar state as the key issue is that development of AI in “decent” sim racing titles effectively stalled a decade or so again as they were just incidental to the real racing multiplayer or a gameplay accessory in console titles. The realization that people (even “proper” sim racers) either like to play offline or that there are never enough decent drivers to scatter across a zillion combinations of tracks and cars has only recently come through in recent titles and they are all in a fairly infantile state of development.

    This stuff takes time, a lot of time and it is good to see it is back on the agenda and competition exists. There are more developers actively involved now than ever, so it probably is the dawn of a golden age, we just aren’t there yet. Or we can think all sim racing developers are evil if we want…

    Liked by 2 people

  8. Developing a game costs a good amount of money and forces decissions regarding how to allocate resources i.e. you have to set priorities. And as of today you have two things that always seem to take precedence for producers because they are the best guarantees (Or so they tend to think) of recovering the investment and making profit: 1) Multiplayer, 2) Graphics. Number one is obvious, if you ain’t in the internet today you are dead in short time, so why spend resources on something as difficult to get right as a decent AI when you have fully developed humans for free all around? The second one again is obvious because no matter what the hardcore crowd says, they are niches compared to all the kids seeking a quick thrill. The kids that will get that game and go online won’t care about hardcore physics, but they will sure want nice graphics. Problem is, great graphics in modern PCs mean 3d models with zillions of polygons, big, high detailed textures etc, which are again expensive things, usually employing external specialists. The recent SCEx fiasco with the V8s was apparently mainly due to problems getting the models to progress well enough, which is quite telling. And how many “massive” mods like those of years ago have you seen lately? Most mods today involve single cars because of the sheer difficulty of doing one highly detailed car.

    So, it is my belief that games from the past (Not just racing games, mind you) had many times great AI because nobody really had to spend money on multiplayer or fancy graphics. Nowadays they need that, and no matter what developers promise the community, in the end they have families to feed and want to sell their game to people outside that niche.


    1. Hardcore sim racers will swear blind that multiplayer is the only form of sim racing worthwhile so why would developers bother? Only to find they get panned by supposed enthusiasts on a site like this. Funny old world…


  9. Even old ones aren’t unaffected from AI problems:

    Tangent: I might create a prc.txt Twitter account (think @fanfiction_txt or @r9k_txt, although I think it will end up like @ancient_fgc_txt instead) that collects quotes from comments (and maybe the posts) from here.


  10. Racing game AI has always sucked unless you were a fan of the Crammond sims. Maybe there are more sims now where it’s overtly broken, which could be related to the increased complexity of software development, but it’s always sucked in general.


    1. but iRacing is suck too..have u watch iR v8SC series? its brokem..not much aggression n close racing in there..no wonder iR world championship is about F1..that fragile spacing racing.


  11. I think Pcars AI is one of only that took a step forward with this gens consoles.

    They are not static and they race sort of like people.
    Because of their AI slots. They recognize how they are raced and respond accordingly.
    It’s also the first AI to make me work hard in a sim.
    I’ve heard rfactors might be better. But I don’t play rfactor, nor do I want too.
    PC and expensive because online not included . Nah


Ratio of vowels to consonants will be monitored. Post at your own discretion.

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s