What this game really needs more than anything else...

Yeah, that’s what happens when I start getting into confusing topics. I might look at the video soon, but what you said is what I meant. I think that is a great idea, but how would the AI practice with modded maps, or is the AI simply having a drivers test before being released to the world?

 

By ramming the wall enough many times until it has mapped all the walls and found out best working way.

 

So pretty much, the AI would be best on Developer-Maps because that’s where they learned mostly, and they would have to practice modded maps with the logic they already know?

 

Basically it learns what is right and wrong given different input "stimuli"

The AI doesn't so much "learn" the environment so much as it "learns" how to react to different input stimuli (like visual input, vehicle angle, speed, ect, ect, ect). When we are driving a car fast into a turn and the rear end starts to come out on us, we know that the proper way to deal with that situation would be to countersteer just enough, and apply light throttle (depending on the car, your mileage may very). An inexperienced driver knows nothing of this however. They will usually slam the breaks with way to much countersteer causing a lose of control. Same with an inexperienced AI and an experienced one. Given the proper inputs, the AI can be taught how to handle the car in just about any way you can imagine.

For instance...

(it's not perfect, but it's close enough)

So essentially, you can train the AI to recognize roads, recognize turns, recognize road hazards, recognize other drivers and potential obstacles, teach it how to deal with all of them, and then let it get really good at what it needs to do with almost any car.

 

I think that would be a bad idea, to require the AI to practice driving on a new map before it drives competently. The AI needs to have good driving skills to begin with, like a seasoned player who's playing a new map, and AI learning won't do that. Personally I wouldn't throw out the entire AI that BeamNG has, my preferred approach and I think the devs would concur is to build upon it. They have a good foundation, it just needs some tweaks and updates.

 

Yeah, that's it. It's such a big process to change, that it would be pretty difficult to overhaul the AI to incorporate that. With what's happening so far with the queueing AI, I bet we're still close. Even if, AI isn't the most important thing in the game, so I'm content with the AI we have now.

 

I would like to point out that I am not saying it is the only solution... just one possible one. And as far as modded maps are concerned, unless they were really outside the learned AI's comfort zone (like... Rainbow Road for instance) I think it would actually do pretty well. Granted I can't know until somebody (or I but I highly doubt I could figure it out) would try it. But I feel if the training data consisted of many different tracks it would probably learn to be fairly versatile. But it is true that this may not be the best way to go about this.

Anyways... I do agree that the AI needs some improvement... one way that might help is to let the AI know just what kind of car it is driving... as right now it seems to be fairly fine as long as it has some really grippy tires and good breaks in corners, but if you take those away from it, it tends to start wandering wide and having unnecessary accidents... That along with what you mentioned before with it seeming to be overly cautious on really wide roads. Normally you would think this is where it would unleash everything it has but more often than not it just ends up holding back.

That all being said though, I am sure the devs are more than aware of all of this and are probably hot on it. Just because they haven't said anything about it in a while doesn't mean they aren't working on it.

 

The last part is what I wasn't so sure about (that the devs are more than aware of the AI issues and working on them). I had the opposite impression, given the very few AI updates we've had over the past year compared to other types of updates, hence I started this thread.

I noticed the AI generally, not always, does better with newer vehicles that have ESC, and it's true that even as a human driver you can get more careless with these vehicles compared to the older ones. So one worthy update would be to make the AI more cautious of throttle-induced oversteer in the older vehicles, but at the same time make the AI more aggressive on wide open roads.

 

Pretty sure most of the community knows that the team can be a bit transparent about updates and development. We haven't heard much else about the Traffic AI, and I'm pretty sure that the Script AI thread came out of the blue for most people.

 

Indeed. The team is cautious with releasing their progress because things change. Sometimes you may be trying to implement something and it just doesn't work out so it is scraped, sometimes other problems come up and that feature just dies as it is left dormant, sometimes the lead on that project leaves and then no one can pick up the slack on it... all sorts of issues. So most times... to keep people from holding them to EVERY new thing they have ever tried, they tend to keep things under wraps until they are sure that it will be added into the game. It seems like they aren't doing much from the outside, but in reality there is about 95% of work that we, the users, will never see.

With an entirely different approach to almost everything you can imagine, I am sure they are trying all sorts of things to improve the AI. The thing is, look at self driving cars right now... most of them are just "ok" at this point (they are improving, but still). They have hundreds of programers behind them pushing their code forward and trying to get the cars to drive as best as possible. Due to the way Beam is made, they are trying to do the exact same thing... except with only about 20 programmer's in total. That and I can guarantee you not all of them are programing for the AI... I would guess about 5 are working on that project. So it will take some time for them to get it to where they want it.

 

I know that AI is on Estama's list and you can't find such skilled and capable individual like Estama from every street corner, not even from every high tech company, but that i thing with BeamNG, they all are incredibly talented individuals, definitely not your average game devs.

AI is an area where they are moving to something that is probably going to create completely new original research, there has been reasons why AI has not been moving on as explained by linked post and other posts in that thread, but as now it is area they can be working, at least I don't have any doubts of it.

You might not see anything about it very soon though as there is lot of work to be done before things are at point where they can start implementing new creations into game, but based on what has been shared on different threads it seems big things being worked at 2019 will be career mode and AI, also some performance optimizations, so definitely AI is on list for this year.

Oh and I think only Estama is working on AI right now, I think atv_123 numbers of personnel doing programming might be bit high, 20 or so might be total number of full time people, but idk of such stuff much, don't really care too much as I know they know their job and they are very capable.

 

No! I like to laugh at the drunk drivers (AI) as they fly into a tight corner at 95 MPH and crash.