Introduction As far as I understand Beamng uses for its AI fully modeled cars with the same physics which apply to the player. In contrast most of the racing games use simplified models especially when it comes to the physics. You'll often notice this when the oppenents pull of actions which are impossible for yourself to do because they are simply not driving on the same physics. Opponents taking corners with 200MPH in NFS games as an example ( rubberbanding is also a part of this, but even with rubberbanding they shouldnt be able to take corners at that speed ). Also Beamngs AI uses the same controller inputs as we(the players) do. Which is pretty neat considering they needed to implemented technology which is also used in autonomous driving to pull this off. My problem with opponents in the standard racing game is that they all behave the same. They share the same abbilities for corner taking, their aggresivness is on the same level, they all take the same amount of risk for overtaking, etc. There were games which tried to randomize the abbilities and preferences of AI-drivers but they havent varied to much and also this games are rare. My Idea is to add different flavours of AI behavior by read out the inputs of the players when driving. Over time there could be data gathered like: -how often do you crash and where -which racingline are you taking mostly on tracks -how consistent of a driver you are -are you more a cruiser, racer or crasher -in which corners are you fast in which are you slow -How you treat your wheel and pedals. etc. On that data beamng could make changes on the parameters in an AI-file. With an option to turn on/off and select an AI-file for data gathering you could also "train" more than one opponent. A trained AI-file which mimics your controller inputs could be used for yourself or passed on to friends, to people on the forums etc. The selftrained and downloaded AI-Files could be : -a good clean circuit racer -a mediocre circuit racer -good at rallying etc. So, i know i just touched on the holy grail of AI-behavior in games. There sure is a reason nobody pulled this off so far. Probably because its complicated like hell. Before this post i looked into racing game AI for an hour and even the simplest stuff bamboozled my brain. But since AI and the players use the same input i see slight chance. Edit: I just remembered Real Racing 3 and the more modern Forzas have something similar in place.
You mean an adaptive AI behaviour? that would take an awful lot of time but could probably be better in the long run. Nice done suggestion, by the way.
There is some variation in the AI when you change the risk level using the AI app. It isn't very much variation, but it's there. I hope your suggestion gets added, but I doubt that they will reach that point of AI development for a few years.
