Better AI! I feel the developers have their priorities misplaced. I mean the new Italy map and car are nice, but what can you really do with them other than just drive by yourself in another lifeless world with a not so great frame rate? As someone who has spent over 300 hours on this game I feel the lack of a decent AI is the biggest roadblock to further enjoyment. I'd take better AI over the upcoming career mode any day. The AI especially needs to take the width and surface type of the road into account to drive at a good speed without crashing or spinning out. I often see the AI drive either too cautiously on wide turns or spinning out because of over-correcting. Please work on improving that.
Something there kinda feels not so well though out. There has been ton of threads where ton of information has provided, also about AI development, but somehow I don't feel like digging them up, something way that quoted part was written is kinda offensive.
You do realise there are different people working on different aspects of the game right? Vehicle and map designers do not program the AI. The AI is being worked on simultaneously to vehicles and maps, just because AI improvements didn't come in this update it doesn't mean it's not being worked on.
Actually AI in Rocky Start campaign is pretty good. I guess its because it goes by predefined route and doesnt generate its way in real time. I couldnt win the second race against Roamer for a long time until I ehm... "helped" him to loose I am really interested in how devs made the AI in that campaign so capable of. Usually during both races Roamer doesnt crash while rushing through rough terrain. And even if it crashes it recovers pretty quickly and even catches up if you drive slower than at the beginning of the race. I dont think its rubberbend like in arcade races so I guess its all about AI capabilities. Or maybe its me who sucks at driving
Exactly what he said. There's quite a bit of news about AI improvements in the micro-blogs. The map developers are separate from the AI developers, and I'm sure plenty of work has been going on beside Italy that we haven't seen.
Game development in nutshell: Research, studying, developing tools, fixing bugs. Then some rather small amount of time for making content and actual development of actual game. It is like you have to figure out and build everything needed to build foundation and then on top of that it is quite easy to build that house, but it takes lot of time to get to that stage. Of course I'm not a developer or anything like that, but I have always been interested in such and have read quite bit of stories from different developers how they came up with the idea and how work did progress etc. Also being in modding stuff for as long as modding has existed I have been around in some projects that have offered window to how work progresses and what are things that take time etc. Also I have very little, but little experience of programming my own tools and some crappy Unity game that I never released, but all that helps to see the process bit better. There still is a lot of it that I don't know, but so much I have learned that it is huge amount of work that goes for some simple things, AI is not simple thing, what I understand is that they want to pioneer something new, which might even require some original research, so it can take time, but I remember Estama mentioning that it is next on his list, that was on where we are going topic I think. Destination is nothing, it's all about the traveling to get there, that is the most fun part.
Putting my 2 cents in here... seeing as this game has practically been built from the ground up from scratch, it is no wonder why things take as long as they do. Most games just start with an engine, add some assets and a storyline, then boom... you have a game. (massively simplified obviously) BeamNG on the other hand basically started with an idea. Then created that idea from the ground up through hard work and massive amounts of programing. So there AI will be fundamentally different from most games not because they want it to be different, but because it basically HAS to be different. If it is not different, then it just won't function properly. It will spin and it will crash way too often for it to be any fun. Most games (and by that I mean damn near all) give the AI little helps that basically allows it to cheat to make it competitive. If we take Forza for example, when you change the difficulty setting, what your basically doing is changing how much extra power the AI's cars have compared to what they are supposed to have. For all I know, it might even increase the grip too... but this is all to make it competitive... and I would consider Forza's AI to be halfway decent at least. BeamNG's AI won't be done like that... I can almost guarantee it because the Dev's just won't allow that kind of cheatery to happen probably... Maybe. Anyways, what the Devs are going for With Beam's AI is basically the same kind of AI that is used in self driving cars. They want the AI to have all the same inputs as a real driver that would be sitting behind the wheel of a real car and for that AI to be able to control that car on the limit as best as possible. That is ONE HELL of a tall order seeing as I have only ever seen one car do that in real life... that would be that one crazy BMW that did a flying lap of the Top Gear track by itself at full tilt. That is what Beam aims to create. Its gonna take a LONG time to pull something like that off... and honestly... would be amazing even if they can just get close.
Very well said. Most racing games such as Forza have a pre-set AI path. But as you said, Beam aims toward realism. Complete randomness. Even when its not fully done, it's still a lot of fun to mess around with.
I'm a bit surprised by the tone of the replies. I think you guys are much too defensive. FYI I love this game as much as any of you and plenty of others have mentioned the inadequacies of the AI. I have no illusions about the difficulties of programming AI. I'm a game programmer myself and have done vehicle AI in the past. It's not easy, but there's always room for improvement. But over the year or so that I've played BeamNG I've seen very little updates on the AI. All I'm saying is that I wish the devs would make it a priority, especially since this game has no multiplayer and the AI is pretty much all we have to race / play against.
Different wording might help as well as if you would look up what has been posted before about subject: https://www.beamng.com/threads/where-are-we-going.59239/page-10#post-961387
OK, thanks for the info, though it doesn't really help that much. It still sounds like the AI isn't a top priority, though it's good to hear that it's not forgotten. I definitely look forward to AI updates, more than any other updates. Like I said the AI needs to better take into account the width of a road. It actually does really well on small roads. It follows the road nice & tight. But on wider roads it doesn't take advantage of the extra space around it, so it's easy for a human driver to blast past it. It seems to me that the AI nodes need a road width and maybe surface type parameter so the AI is more aware of its environment. And also I'd like the AI to be more able to get out of oversteer.
Alright, fair enough. Despite what I said earlier, I do agree. Granted, given the complexities of operating a car at the limit like is needed, I am wondering if a different approach should be taken. Like... for driving around town and what not... your never really on any sort of limit, so the AI when driving like a normal person can be pretty simplistic while still being fully functional. But when you start pushing a car you almost need a completely different mindset... now you have to start monitoring grip levels, start monitoring forces, looking way out ahead of you at the next curve in the road, scanning the road for puddles, potholes, or things of that nature. I wonder if for fast driving... if the AI should be set up like a deep learning neural network. I mean... think about it. It might be right up BeamNG's ally. It is REALLY hard to take the inputs that are available from driving a car and to just code a system that can drive at the limit with those few inputs... but as we keep seeing again and again and again... deep learning neural networks can essentially learn how to be really good. Plus, once it is built up, the neural network's weights and biases can basically be dropped into a single equation capable of now controlling a car. More research would need to be done in this area naturally (heck, I don't even know all of what data you would have access to from Beam to even try such a thing) but I feel like it could be done. For anyone that doesn't know what I am talking about... Yes this is with an arcade game, but the same ideas apply.
First make sensors so AI car can see, knows what is steering, can be told what are limits, I mean they want to simulate it, not just to do it simple fake way, so for example radar, IR and so on might be needed to be simulated first etc. I don't know how far they are going with it though, but no cheats approach is quite different than typical AI.
I think all of that can actually be obtained interestingly... take a look at this (I think it is WIP, but still impressive)
I know, that is why I did mentioned those things, I'm not sure how much of tech will be in game side, but it is clear to me that no cheating AI will need some of such kind or then it would be bit of cheating. I guess that for game some lite/light version is needed.
Agreed... plus for the game AI, drawing in that much data for each car would be really really limiting in terms of how many AI could run simultaneously. Personally I think that everything the AI can see should come from the standard interior camera view... you know... so it acts more like a real driver... plus that would cut down on the amount of data it could see saving computations.
Well you guys are getting ahead of what I was actually asking. Making the AI "see" would of course be nice, provided it's optimized for the game, but that's beyond what I had in mind. As a software engineer I've always preferred simpler solutions to more complex ones, if the simpler solution is sufficient, for example putting an invisible "stick" on the front of an AI car and checking for collision with the collision box of another car when deciding whether to stop or slow down so the AI doesn't hit the car in front of it. Depending on what the BeamNG dev team is after, this may not be a good enough solution for them, but I say it's better than none (currently BeamNG AI doesn't care what's in front of it). Likewise to implement more aggressive driving on wide roads, something like what I mentioned above could be implemented, even if it's always work in progress. I just want to see progress, and the AI in this game could sure use some.
As a BeamNG AI aficionado (not), I'd say the whole 'Learning' thing is the best thing the AI could do. Remember the Herbie mod and how it remembers just how Herbie feels? Maybe something like that could be used, along with the whole 'Learning stuff (I watched a Code Bullet video about AI learning, still confused). For instance, the first time the player downloads BeamNG, the AI would be pretty stupid, hopefully on the level the AI can drive now. The more the AI is used, the more the AI "Learns" about driving and the whole system. Soon enough, the AI can race around Italy like it's a sunday drive. Of course, this would mean a complete overhall of the AI system, so that would probably not happen. All I know, I bet one of the devs have some "AI Updates" folder somewhere that will go along with the Career Mode update. I mean, anyone else remember this thread? https://www.beamng.com/threads/ai-work.54961/ This looks amazing for the game. Once the AI is able to 'see' other vehicles on the road, then maybe they will have an easier time with most other things. The only supposed drawback with that is the AI might only be able to read "Autojunction" paths and queue for it like that. Either that, or they would have found some way. About the whole "Easier way" thing. Knowing how long this game has taken to reach even this point, the devs don't like to cut any decent corners and release it. With how complicated everything is, it'd be a blow to everyone if they did that. Also, we got Script AI, so that's one way to set up scenes until AI is able to buy glasses. Honestly, the AI in this game is the only thing I look forward to playing with, so I'd be more then happy with the AI being amazing, but this is still...kind of groundbreaking stuff, so it's not surprising that it's taking a while. You said you're a developer, so I guess you can understand just how difficult this stuff is to do, especially with rarely having a chance to really drive fast and feel what a human racing along fast would feel. They said this a while back, but at the end of the day, the AI is a version of how smart the developers are. Even if, there are more important things in the game that need to be fixed, and even I can wait patiently for that...I'll just lay on my bed and wait for that one changelog patiently with a smile on my face because of the game we have now.
That wasn't quite where I was going with that, but close enough I suppose. I have been rather heavily invested in Neural Networks and Genetic Algorithms over the past year to year and a half. I could try and explain everything to you... but honestly I could never do as good a job as this video does, so I suggest just watching this. These networks, interestingly enough can "learn" to do damn near anything, and even though there is no actual "intelligence" in the system, they can learn to do it really fast and really, REALLY well. Plus, they use very little computing power once they are done learning. They are nothing more than a big mathematical equation that (to our minds anyways) seems to make no sense, but to the system, is a fully functioning AI capable of doing exactly what it was trained to do. Now while you could train it to do city driving... I personally would train it to do the racing. I feel like you could program the city, low speed driving rather painlessly since (like I stated before) your never really on the edge of control of what the car can do... your just driving around. But to program a system to handle a car in the way that Beam requires is another task entirely. That is where I think a system like this would shine. Granted it would take a while to learn... like... I am talking potentially about a month to actually learn how to drive properly, and then another to actually get good at it... but once your there, boom... your done. Just slap the final result into the game and it's basically a done deal.
