Could Ai Speed Up The Creation Of Route Builds?

With the constant presence of AI and its evolution in the last 15 years, we’ve seen a shift in big corporations introducing AI into their day to day business practices.
AI isn’t going anywhere, it’s the backbone of the future we’re heading into, whether we like it or not.

We’ve already seen it in technology used to upscale, improve fps in gaming with NVIDIA’S DLSS (deep learning super sampling) and PlayStation’s PSSR (PlayStation spectral super resolution).

What if DTG started using AI to assist in route building? It would build the world environment and world assets a lot quicker for starters, cut down development time.
At some point though human invention would have to occur, with tweaks to correct/improve what the AI had created.
I don’t think AI is at that level currently for such a radical change in game development, but it’s on the horizon for sure.

 

I'm presuming some sort of AI was MS/Aces reasoning behind their World Of Rails idea for their second attempt at MSTS2, which as we know went nowhere. In the 25 years or so we have enjoyed fully graphical train sims, no one has yet come up with a better way of building the routes than plonking items down longhand so to speak in a graphical editor window.

 

We don’t need clackers (ai bots) working on the game

 

Procedurally generated maps have existed for many years already

 

That’s how it should be. It keeps people in a job, also human created ideas look better than Ai slop. You’ve only got to look at corporations like Activision when they tried to fool their fan base with Ai generated assets (calling cards). The backlash they received was well deserved. It’s easy to tell Ai created versus human created.

 

With all the backlash AI companies keep getting, DTG would be smart to steer clear. Unless of course ChatGPT knows how to fix blurry textures

 

From personal experience, it will certainly act like an expert on the matter and believe its correct 100% of the time.

 

Will fit in in this forum then

 

Yes it would probably just delete all textures and declare the issue fixed

 

It can definitely be used as a tool to gather the data they need to build. But from how it’s been described to me from my more detailed tech friends, upgrades in the UE tech with a move to UE5 will help much more with route building. It’s not foolproof and brings its own set of challenges, but the asset workflow part alone sounds like it would be a major jump in either speed of build or size of the build within the current timeframe.

 

An AI model should be built and used for complex signalling and pathing issues. All of this ‘Stop off Path’ or the train jams we’ve been seeing could have been mitigated easily by a smarter signal/pathing ai model.

It would be a great evolution of the simulation to see a dynamic railway unfolding around you - with delayed ai having to utilise alternate platforms, bi-directional lines being utilised effectively (to stem opposite flows of traffic) and even different driving profiles of ai trains (weather differences / relation to timeliness / head code based).

 

People who say AI is going nowhere really underestimate what’s happening. We’ve seen this before with that “little thing” called the internet—25 years ago, plenty of people said that was going nowhere too. Funny how that turned out.
The truth is, we don’t yet know exactly where AI will lead or how big its impact will be. What is clear, though, is that it’s a huge deal. When the world’s largest companies are investing massively in it—and when global power dynamics, energy concerns, and resources are being reshaped around the race for better AI—it’s hard to deny its importance. Anyone dismissing this entirely is probably missing the bigger picture.

Will AI eventually build routes for train simulators? Very likely. But it probably won’t be based on Train Sim Classic or Train Sim World as they exist today. It’ll more likely be something entirely new. The real question isn’t if, but when. And if we take the major AI companies at their word, this could be possible within the next two to three years.

 

You’ve only got to look at how MSFS uses real world data to map the surroundings around you in real time. Nothing stopping other simulations adapting these practices. As time passes it’s only going to evolve and improve.

 

A lot of this stem from the fact that people don’t actually know what AI does. They assume Skynet kind of stuff instead of calling it what it is, a data hyper-processor.

Many people think it’s the golden source of information without realizing that it’s still being given information from humans. I always laugh when people say they asked ChatGPT something and react like the response given isn’t something that is just a compiled list of human provided inputs (not to mention the lack of accuracy in some cases due to bad data).

 

Maybe it could do something, but could it do route building better? Absolutely not. Could it perhaps be used to make 3D models? That I could maybe see, but I doubt it could make detailed close in models in the same way a talented modeller could.

Think about it, with all the might of Microsoft behind it the AI generation used in Flight Sim is amazing at what it does, but it only works because it’s a flight sim, and it only works on generalised distant scenery. The AI generation of airfields is absolute rubbish, and all the up close and personal scenery needs to be handmade by humans. It’s why there’s such a vibrant market for scenery DLC.

By contrast, by its nature the whole of TSW is up close and personal, and only a human can make it look good. Route building (like modelling) is a bit of an art form, and it is not something a machine can ever replicate. I’d put money on it not being in my lifetime that AI could even get close. Some other new fad will have come along by then that will be the new big thing and AI as we know it today will be consigned to the history books.

My personal opinion is that AI is massively over-hyped anyway. It has its uses, but almost everything I’ve ever used it for has just been a bit rubbish.

 

As well as gaslighting you into thinking its right and you're wrong. Or outright deny there's anything wrong to start with.

 

100%, lot of airports have issues in MSFS by default, like I was at the Clinton's airport in Arkansas once, one of the taxi ways was supposed to have a bridge, but the gen instead just had the taxiway dip into the area instead. Remember it vividly because I was taxiing there finishing up a night flight, I luckily spotted the dip just before driving straight into it. I don't even remember how, because distinctly one thing about those autogen airports is that they often had abysmal night lighting too. Don't get me wrong the tech's still impressive, but for pretty much any other game this tech really isn't all that handy.

Really MSFS also has insane world requirements for a video game, even the biggest game worlds in IRL terms barely get bigger than a moderately sized city. Like GTAV's world feels massive, but I think the whole area in reality is around the size of just Denver International or something like that, nothing close to the actual scale of a city. And while it's old-ish, I don't think there's a ton of open world games bigger than it, and arguably even the ones that do exist And even at those scales there's tons debate about how empty certain games are. I remember Twilight Princess getting a lot of flack in select circles for its empty Hyrule Field back in the day, and that was for a frankly tiny area by modern standards.

 

Well open ai has openly admitted that they will go bankrupt without more funding, it’s already wasted 14 billion so honestly don’t think it’s going anywhere

it’s just cheaper using humans over fixing ai’s mistakes

 

Artificial intelligence could help in creating new routes, by correctly placing signs along the railway line. For example, deploying PZB magnets. I would like a dispatcher mode, where all traffic is taken over by a trained artificial intelligence. The dispatcher could then learn the possibility of overtaking a delayed train, using other platforms, controlling the train's shift. There are many such possibilities. Moreover, a well-written AI model can learn and improve.

 

hi
Here is the new AI generated route. You will notice that at York station there is half of Eiffel tower because the AI data scraped youtube and included a blog about someone's train travels to Paris. There are also vikings as passengers

 

Nobody said that, especially not in the early 2000s after it had been well established for over a decade.
But if we want to go back to things which emerged out of the 1980s, the AI boom back then also failed to deliver on promises and went bust. Basically killing investment in the field post-1987, until investors got desperate for any moonshot gamble that might possibly get them a financial return post-2008.

A better comparison would be hydrogen cell vehicles.

Similarly got massive, in this case Europe-wide investment of hundreds of billions in developing the infrastructure, and just like we're seeing with the explosion now in data centre cancellations, it simply was not sustainable as a cost-effective solution for anything it was being touted for. Producing hydrogen resulted in everything being more expensive than just putting the raw electricity generated directly into the grid and charging battery electric cars. After a huge surge in sales of hydrogen cars in Denmark and a deal with Germany to build a pipeline to supply hydrogen internationally, the main producer of it in Denmark had to scale back, pull out of the international pipeline deal, because it could not cost-effectively scale and was turning into a gigantic money sink, public facing charging stations were closed across Denmark, because they could not produce it even as they got more and more money, leading to ongoing legal action from owners of hydrogen cars which can now no longer refuel in the country, and are essentially the world's biggest paperweights. Then the country's biggest producer of hydrogen went bankrupt. The state pulled all the funding, realising it was mathematically impossible to produce it any more efficiently, and the industry there collapsed. Similarly, just this week gone, Norway's main hydrogen refuelling supplier went out of business, after only NINE cars were sold in the country in the last two years.

The best advice I can give for the next few years when it comes to the data centre projects which HAVE been built but cannot be turned on due to lacking electricity capacity and are sitting there empty with the warehoused GPUs losing value by the day, is invest in Laser Tag.

 

I don't think it is, it's not even beyond the horizont. The issues people experience when working with LLM (overconfidence, hallucinations) are not something that can be easily ironed out, as they are rooted in how LLMs operate. They are not smart, they don't really understand context, or nuances... they are just extremely large statisticall models... They are good at creating drafts, but they cannot create a final product in acceptable quality without supervision. Which is something the companies who fired devs, becasue of AI bubble, are quickly realizing.

From personal experience - the more complex code you ask LLM to generate, the more time you then spent fixing all the issues and clean code violations the LLM result produces. For larger stuff it quickly becomes non-efficient, just by looking on how expensive developer time is, without even factoring how expensive running those models is.

They are usefull, but I don't think that they present any radical shift in game development. It is just the next generation of automated tools whose biggest benefit is time and cost reduction of mudane tasks. I don't think they will take over the specialized tasks, like signage placement (becasue you don't want to place them any other way, than what they are in reality), or modelling important stations or landmarks. But in generating 'filler' models of cities/towns, foliages, adding realistict grime patterns... etc. Tasks that are repeative, and not under that much scrutiny, and you have a statistic set where to train the model on. You cannot train it to build Hamburg Hauptbahnhof, because that is a one unique building in the world. But you can train it how a typical german small town 2-storey house looks like, and it can generate 100 of them.

And believe me, no artist really wants to model 100 variants of the same thing, thats why those tasks are usually done differently even now - i.e. you split the model into parts, do variants of those parts, and then do permutation of those parts to get variety.

 

Winzarten Yes, many people are saying these kind of things. Thanks for insight. As of today, AI is not replacing everyone. But will it stay it like that? It´s been said, it´s only a matter of time. Maybe it will take longer than we think. Or is it all just a massive hoax and complicated task will never get replaced by AI? Who knows... I would say that there is something to it, when everyone is pumping money towards it and China is also chasing it.

 

I'd say that AI should be used for creating distant scenery, just like MSFS. With those resources saved the route builders could focus on more details near the track. Hope it's better than the tech they used on LGV...

 

CEOs think it can do anything, because it can automate the 50% of their job that is not going to brunch (writing vapid emails).

 

It's been mainly said by marketing people who want to attract investment money. And that's the issue with the 'AI' label that gets thrown around - it is a marketing label... not a technology label, it's a buzzword... like 'smart' was before. The technology is LLM, GPT (Generative pre-trained transformer)...etc

The technology isn't that new (I think IBM had some LLMs in the 90s), it's just that now they got some use for the general public, mainly because of availability of cheap (*cough* pirated *cough*) training sets, and cheap + powerfull hardware, that allow models with billions and trillions parameters. The limitations of LLM are fairly uderstood.

Noone can predict the future, and there might be a technology that will also be able to replace the creative and cognitive aspects of human brains, but it won't be LLMs. LLMs can still be usefull though, and can do things in ways that weren't possible before... but they are not the singularity type event some people make them to be ;-)

 

bold to assume AI would be even remotely good in such complex fields like safety systems in a simulator when it cant even make pictures of trains that dont have tons of inaccuracys. What i have to see in the internet these days with AI slop of everything hurts

 

That’s what I’d like to see Ai used for too. The asset building needs to be left to man made creations. We’ve already seen enough Ai generated slop in today’s gaming.

It would save a lot of development time if they could cut 50% of the work load. Would mean quicker releases for everyone.

 

Frankly, when it comes to TSW, I'd rather see them focus on applying ML systems to improving game performance by implementing stuff like DLSS, and leave the creative work to the existing teams. there are enough corners being cut already to meet deadlines. Maybe instead of asking where generative stuff can introduce potentially even more glaring inaccuracies in need of fixing post-release, potentially generating more work, they should look at the proven applications.

 

Yes, that’s where Ai shines currently (DLSS) improving picture and fps quality. I certainly notice that a lot when playing PS5 Pro enhanced games, okay that’s PSSR, but it follows the same principles.

Performance is the number one concern for most players, so at some point if DTG do start using DLSS and an advanced version of PSSR for the PS6 than it would be beneficial to the company. I say the PS6 as I feel PSSR for the PS5 is nowhere near the level that it should be.
The team can concentrate more on the route building. I can imagine a huge chuck of development time goes on optimisation, would be nice to cut that down.

 

You can't compare image generation to automatic train control. If an AI model learns to control train traffic, so can the regulations. Development is moving very quickly. Proof of this is how quickly mobile phones have developed.

 

You cannot just cite at random something unrelated as "proof" something else is getting better. That's like me talking about how the rapid pace of development in catalytic converters through the 1980s is "proof" of how successful Betamax is going to be.

 

A company like DTG can not afford to have a LLM trained to do so. I have the feeling some people here getting the "AI idea" in the wrong throat. Same with the DLSS stuff. At the very first, DLSS and the other things around are simple upscalers. Nothing fancy new. Working somewhat in TV sets or sat-receivers for two decades nearly now. You need to differentiate between the on-the-fly up-scaling and the trained ML upscalers. The latter is a massively expensive thing to do. For a game, like TSW is, the non ML DLSS does nothing much. Especially not making the image better or the performance smoother. All it can do to TSW is inserting not existing data to double the framerate visually but it will not gain any performance at all on better systems at the same rendered resolution. All it is good for, and what it is used for mostly (even on PS5) is to render a lower resolution (what gains the performance) and upscale the result back to the set screen resolution. That process is coming with a bunch of problems like artefacts, input delays and such. ML upscaling is better in these things because it gets the game to eat beforehand and extracts a algorithm out of that. That might work on a game with a single not changing world or some levels that are always the same and not too big. TSW constantly gets new maps and content the algorithm doesn't know about then and needs to re-learn all the things again (for each route basically because the route is the world in the game). And routes in TSW are way bigger in terms of world size (the overall landscape size the ML needs to eat) than any other AAA game that might can afford ML DLSS algorithms.

 

Where I could see AI possibly being useful is, say, scanning a defined map block for railroad tracks, and doing a quick first pass rendering of selected tracks and scenery. Nothing too detailed, but it could remove some of the initial tedious grunt work, letting the humans focus on details and polish.

 

I mean, I'm still not convinced.

Even when it had the direct source material to work with

One of these models still managed to LOVE out this

 

whilst i don’t think an ai-created route would ever be a good idea, i still think there is something to be said for ai. smaller tasks such as identifying a bug, improved announcements, and guard mode would all benefit from ai use i feel.

 

Oh, no, I didn't mean that. "Map" was definitely the wrong word to use. I meant an overhead satellite view like Google Earth or such, something where you can see the tracks. AI can scan the picture and pull out what we want much faster than we can do it by eye and mouse.

 

Except it literally scanned a map here, an EXTREMELY simplified one, and then spat out a wholly inaccurate recreation of it with a bunch of additions. I don't see that scaling well to something more complex. We already HAVE good maps of rail infrastructure, surveys of topography, etc. to copy as a basis to build track layouts from.

 

I find that sometimes it takes a few attempts even with the best model to get what you actually want.

Other times it just can't read the provided image so just guesses.

 

Enjoying this AI discussion in this context.

 

Good thread.

Most likely outcome with AI in the short/medium term, is an awful lot of people losing an awful lot of money.

 

You mean those that will lose their job ?

 

Oh dear, I wish I could permanently live in a time "pre-AI", we are in for a rough time with this nonsense..

 

Couldn’t agree more. You ought to see how poor so much of the google maps coverage is/was around MVL. We had a real job interpreting it, a machine would have absolutely no chance!

 

I use Gemini. It helps me with Excel. Last week Gemini helped us fix a broken engine on a customer's car. It pointed us to the faulty engine component. It's a useful tool for me.

 

They said this about the PC. They said this about the internet. Jobs weren’t destroyed, they were however changed.

Point is, people will adapt and there is always a need for some sort of human involvement even when tech is involved.

 

This discussion reminds me of the application of an invention called the steam engine at the beginning of the 20th century.

 

Who says they're not already using it?

After all, they're not obligated to say so...
At the same time, given the number of bugs per DLC, it wouldn't surprise me at all...

 

That does tend to be the case with bubbles that collapse the economy when they pop, but annoyingly it's always the conmen who were selling all the monorails who get bailed out.

 

I see AI as the tool that takes care of the beans, so developers can spend their time on the good stuff, the meat and veggies, leading to a faster, better, and tastier Chili con Carne.

So yeah, it could positively benefit route development too!