AMS2 AI Developer Q&A - Physics & AI

Assuming late is indeed better than never, I´ll finally be following up on the above and should have the remaining replies up over the next 24h or so

If anyone has a well substantiated physics or AI enquiry they might want to throw into the mix, this would be the time to do it.

 

Locked for further questions folks, replies incoming - thanks again everyone for participating and especially those who held patiently for an answer all these past weeks

 

Alright folks so to wrap this up at last - working my way back from the last post for convenience, if I end up missing anything from older posts please feel free to give me a nudge on PM about it.

1- By "cheaty" I´m assuming you mean it´s too fast in parts of the track? In any case yes, in fact there is way more than the AI lines currently being overhauled for the Nordschleife so there is going to be a complete redo in upcoming updates.

2- This is not planned nor easily achieved in the foreseeable future.

3- Not currently, there is enough work to do with existing weather & track dynamics for it to be worth throwing a more niche variable that doesn´t really have much of a context within AMS2; with that said on a long enough timeframe anything is possible.

4- No

 

1- There is in fact some existing logic to reduce the AI´s "gravitational pull" towards their predetermined paths during an incident but it´s fairly basic and there are indeed cases when the AI driver´s urge to get back in line while still tangled up with another car will make the incident worse and / or harder for the player to untangle from it; certainly an area we can and intend to make improvements on still.

2- This is trickier to address because AI physics is simplified in several areas and in many classes these simplified physics require stabilizing forces in order for the AI cars to remain stable and competitive vs the player; as touched upon in previous replies the problem can be the opposite in some classes where AI cars still can suffer from instability particularly at higher speeds and in sections involving agressive changes of direction where they can lose control on their own; without these stabilizing forces those issues would be far more widespread and that is a bigger problem than the AI being a bit too robust in retaining control when tagging or being tagged by a human car.

While I´d agree this is still neverthless a con in AI racing of AMS2 with some classes I´m not sure the magnitude is what some users might believe it is; I´ve seen many such video reports supposedly caused by this in which AI tagging player or vice versa which just the dynamics of the contact would have been enough to see the player coming worse off even with everything else being equal.

 

1- I´m fairly sure AI has ABS when it´s meant to but the particulars of which version of ABS it uses (the legacy pMotor version or the Madness system used by the player) and what sort of simplifcations apply is not something I can answer from top of head.

2- Yes to track limit violations, but not as in-depth as it applies to the player (it won´t be forced to slow down and / or give the position

3- AI Skill and Aggression settings have no impact on AI power output or car setup; there is actually an error in one of the UI helping texts that contradicts this information but that is due to that text still being legacy from SMS we have missed out on correcting, in terms of game mechanics we have since very early on development disabled all effects of AI Skill on anything that isn´t strictly related to driving ability and a a driver cannot make a car go faster on a straight irrespective of how skilled he / she might be (assuming of course the car is stable on said straightline); the exception of course would be if said driver came out faster of the previous corner leading into the straight, which might be what you observed in your AI tests at different configurations.

 

Would need to verify whether it is indeed something particular to the track (ie every class is consistently faster in those corners than what user can manage at his / her standard AI Skill setting), the car (in which case the F-USA 2023 AI physics in speedway configuration might need a calibration pass) or a bit of both; either way as with most performance discrepancies it shouldn´t be especially hard to at least mitigate it so we´ll take a stab at that for the next update

 

To add to Marc´s reply above, even if developers didn´t have to contend with this system performance consideration we still need to realise that AI in AMS2 (and in all major sims AFAIK) does not simply drive the car the way a human player does; it is subject to a number of predefined variables, calculations and predictions, the car physics is just part of that equation and while removing that variable completely could help the balancing process in some ways it could also hurt it in others where those remaining AI resources might not be enough for them to handle the cars we have competently.

Ultimately the goal to deliver on with AI bots is for them to a) be able to go around the track safely (whether on their own or surrounded by other cars) and b) with a window of performance that is tangible to what human players can achieve; how one delivers on that is less critical than delivering on it. Our process is time consuming to be sure not least because the range of car / track combos as well as the settings and conditions users can run them under is so vast in AMS2 but the fact progress is constantly being made not only ironing out persisting discrepancies but also creating more paths to address the smaller nuances of driving performance should be evidence we´re not at the ceiling yet.

 

a) Yes this will be introduced eventually; there is in fact a number of valuable developments around AI for MP scenarios we intend to tackle still.

b) This is very unlikely to happen; simply put it would require a lot of development effort for relatively minor value.

c) These sort of behavior dynamics are constantly being adjusted not only at the macro level but also class to class over the course of development; one thing to try if you find that to be a constant problem is to reduce Aggression setting one level as higher aggression settings certainly can make AI less willing both to back down when attemping an overtake as well as concede when being overtaken.

 

Creating new tires for STM unfortunately isn´t something someone without knowledge and in-depth access into all areas of the physics engine could do and it´s even more complex if AI is thrown into the mix.

With that said we´re not against creating value by throwing together some alternative versions of existing cars (or tracks for that matter) when that might make sense - we have done that with oval variants of Brazilian Stock Cars and Super V8s to mention an example and I´m sure we´ll do more in the latter stages of AMS2´s shelf life..

1- Similarly to previous replies on the matter, the AI´s "raceability" is not a constant from class to class and from track to track and generally speaking faster classes still err on the conservative side for the sake of minimizing incidents. There is quite a bit of content-specific fine tuning we can and do perform from update to update to improve their ability to race each other; sometimes these adjustments make a big difference, sometimes not too much but there´s a lot of such tuning still to be done all around; likewise there are also code-side developments constantly being introduced to try eliminate or at least mitigate some of these behavioral "vices", we just need to be careful with those because when it comes to AI it´s far easier to break things than it is to make them better.

2- It´s a valid suggestion, although my first thought is that it might prove tricky to implement in practice as minor driving mistakes can also have grave consequences and vice versa depending on the circunstnaces around the driver making the mistake.

Wind is simulated but is capped to very low speeds so the effect in vehicle performance is generally pretty negligible; that cap might be removed at a latter development stage when aero physics development has settled down.

 

Alright this is a big one so let´s try synthetize replies a bit:

• STT Ai having no visual body roll or limitations with the front end on left-hand corners. There's moments where the player vehicle has significantly less grip than the Ai.

The lack of visible body roll from AI STTs is a bummer and one I hope we can find a solution for over time, but there isn´t one up for grabs atm; the performance discrepancy issues are always viable to resolve or at least mitigate so long as we´re aware of where they persist.

• Kartcross often being unable to cleanly finish a lap.

As with all kart vehicles they are extremely tricky things to get going around the track safely and within a window of performance that is tangible for the player. but this is certainly one problem we can and must fix in the coming updates.

• RX Ai occasionally having formula car levels of downforce and taking "traditional tarmac racing lines" through hairpins and other corners that would require using more slip angle/sliding.

The performance discrepancy around certain corners can certainly be addressed or at least mitigated; as cool as it would be to see them handling the track with a more human-like "body language" and taking more human-like lines that would be both a less tangible and less critical goal.

• All RX classes often fail to take joker laps at certain tracks.

We´re aware of these and still pushing to fix them.

There are several factors contributing to this performance discrepancy but I believe the thick of it would be down to dirt paths across all RX layouts currently sharing the same material with the same properties (grip, resistance, softness etc); this means the dirt sections in tracks like Hockenheim RX are set with the same properties of a track like Ascurra, which is a fairly hardened surface with far less actual dirt as you can see in the video below:

This isn´t a inherent limitation; we could create custom dirt materials for each track and the main reason we didn´t for the dirt content as it stands now is because of some the AI troubles that arise from more intrinsic terrain modelling - things like the AI sinking into the terrain and then "exploding" for ex are aggravated with softer terrain properties.

Other compromises with dirt content include how the tires are impacted by all these surface variables and are likewise mandated by the need to minimise discrepancies in performance vs the AI; these aren´t unsumountable issues but they would require considerable more development input then we could justify at the time of their original release.

We do plan to have another go at dirt / off-road content on the longer term however which should hopefully allow us to solve or at least mitigate some of these design compromises; it might take more time than one would prefer but there is not bit of content in the game that is ever left behind, they are all constantly being revisited and improved as best we can in regular waves of development, take the karting classes for ex. which were in pretty messy shape for a while there and became much better following the updates earlier on this year. It´s only a matter of time until the dirt content gets some love again