Assetto Corsa AI Video and Massarutto Statement

Paul Jeffrey

Premium
Assetto Corsa Multiclass.png

Following the big Assetto Corsa PC update from Kunos Simulazioni, Aristotelis Vasilakos has released a nice video of the new AI code - with some interesting words from Marco Massarutto.


The main thrust of the 1.14 build update released last week focussed around the single player artificial intelligence, an area of the game lacking behind its main sim racing rivals in recent years. Now following considerable work from all involved with Kunos Simulazioni, the current generation build of Assetto Corsa has moved on leaps and bounds since its early days to arguably move very close to the top of the AI pile. With passing and defending from the AI cars and human like behaviour, the new build really is a substantial step forward from the development team.

To mark the new build release and in something of celebration of the new AI physics engineer Aristotelis Vasilakos has released this excellent video of a multiclass AI field racing around the Spanish Barcelona Grand Prix Circuit in Assetto Corsa:

By way of his own personal Facebook page, Marco Massarutto added the following words with regards to the responsibility structure at Kunos:

"I would like to take advantage of this gameplay video to say something. We are a team of crazy, genuine people who love their job, showing some talent but, even more, a genuine devotion for what they do. This makes Assetto Corsa what it is today. But I can't forget to mention that, most of AI&physics algorithms, as well as the tyre model, the netcode and the graphics engine of our game has been programmed by one single person, my teammate in business Stefano "Kunos" Casillo. .

All the tracks (except for one) have been modeled by one guy, Simone Trevisiol, who takes care also about shaders, post processing effects and gfx fine tuning. Most of the sound environment and effects has been managed by one guy (with some programming support), Luca Sodano.

Only one guy, Aristotelis Vasilakos, is behind the car handling, vehicle dynamics and tyre data (that means this lazy guy updated all the hundreds of cars included in the game each time Stefano updated the code behind physics and tyre model). And Gianluca Miragoli is behind the 3D modeling of at least half of the car models and most of the interiors. I should mention all the others that made Assetto Corsa what is today and I thank all of them. But the people mentioned above are with us by the beginning. I'm proud to be part of this team.

Oh, btw: the video shows the new AI algorithms in action: the Toyota TS040 is driven by AI (aggression set at 45%) and it has been placed in the last position in order to show its capability to overtake slower cars at Barcelona circuit , where overtaking sometime is not an easy matter, and it was pretty difficult to see before this update. Sorry to be late, but good things need time.

Assetto Corsa is a racing simulation available to purchase on PC, Xbox One and PlayStation 4.

If you enjoy the new AI then I'm very pleased for you, it is very good indeed. Want to know what's even better though? I'll let you in on a secret... racing against real people, online, in organised events with an emphasis on respectful driving... enter into the equation RaceDepartment Club Racing! We have the best cars and tracks from the game, great event admins and a lively and enthusiastic club of enthusiasts with some awesome teamspeak banter. Suitable for rookies right through to aliens alike, out club events are for you, the sim racing enthusiast. Get yourself involved and sign up for a race today*

*good times guaranteed

Enjoyed the video? Impressed with the new AI? Liking AC post V1.14 release? Let us know in the comments section below!
 
Last edited:
Can the AI pit now? v1.14 a couple more tests.

Tl;dr - the AI race and pit perfectly at Silverstone GP. At the same time, the same cars were a disaster at Nordschleife endurance with only 4 cars finishing the race and half the field did not pit. My thinking is getting a good endurance race from the AI could be quite track dependent at the moment.

Silverstone GP 60 lap endurance race. Hugely impressive performance by the AI. :)

Silverstone GP - 60 lap race, AI settings 100% Strength, 2% Variation, 100% Aggression. Full race weekend including qualifying. A mixed grid of LMP1 and GT2 cars (GT2 as classified by AC) as you'll see in the screenshots. Data from pTracker. As before I retired my car and let the AI get on with it.

This time the AI were much more successful in providing a great race and yes they all pitted as required. There are still a few odd behaviours in the pit strategies, with some cars stopping 5 times and spending over 3 minutes in the pits, and I'll dig into the data a little more below. Only 2 AI cars crashed on lap 1 and LMP1 Porsche and LMP1 Toyota and are not shown in the image.
Assetto%20Corsa%20SilverstoneGPr.png
[/URL][/IMG]
Assetto%20Corsa%20SilverstoneGPr.png

The leading Porsche and top 7 cars pitted twice for fuel with sensible pit strategies such as this:

Winner pit strategy part 1 - for you Aliens who can beat the 100% AI times here they are, and they are consistent:
Assetto%20Corsa%20SilverstoneGPw1.png

Assetto%20Corsa%20SilverstoneGPw2.png

By contrast, the Toyota in 7th place stopped 5 times for fuel using this strategy, stopping every 6/7 laps followed by adding enough fuel to finish the race on lap 36:
Assetto%20Corsa%2025_05_2017%2010_31_51.png

Assetto%20Corsa%2025_05_2017%2010_32_08.png

In the GT 2 class the Porsche 911 RSR 2017 was able to stop just once, it ran in second place in GT2 for most of the race eventually overtaking the Corvette C7R when it made its 4th pitstop.
Assetto%20Corsa%2025_05_2017%2010_44_25.png

Assetto%20Corsa%2025_05_2017%2010_44_35.png


Nordschleife Endurance. A bit of a disaster.:(

Same grid 24 of LMP1 cars mixed with GT2, no qualification this time but cars placed in order LMP1 then GT2 to avoid any possible start line issues. AI at 95% strength, 3% variation and 85% Aggression. I again retired my car to the pit.
Assetto%20Corsa%2024_05_2017%2018_57_34.png

This time 9 cars crashed on lap 1 and as a result are not shown in the image above. The cars in places 6, 8-14 all failed to pit and ran out of fuel. The cars in 5 & 7th crashed out of the race. The Toyota in 5th place that had been well out in front again crashed at Hohenrain against the barriers.

Final thoughts:

All I can say for certain is that the AI works well at Silverstone GP but has issues at Nordschleife Endurance, this could be to do with the lines they run under the new AI but that is just speculation.

If you are thinking of running longer races against the AI it's worth understanding how they behave on the track you chose. It's also worth noting how long the 911 RSR 2017 can run because that car seems to have a significant fuel range advantage and I assume must also be easy on its tyres.
 
Yes, ISI-based sims use "simplified" physics for the AI. Tire model complexity is reduced and suspension is eased to allow the AI cars to use more binary throttle/brake... this allows more complexity, more CPU-time, to be assigned to the player's car. AI still has player's engine & aerodynamics. Grip, power, & brakes are adjusted by the AI Difficulty setting, though one can change that from default if one chooses.

What rF2 has over the other ISI-based sims is multiple racing lines, including wet, blocking, and car-specific lines. rF1 had AI learning and rF2 extends that learning with two more modes.

From a gameplay perspective, it's whether the AI actions are convincing rather than whether they use "player" physics.

Yep, considering that in case of full physics for AI artificial forces or "cheats" are being used so AI can handle those physics I would agree that result is more important than method
Whatever is better we have 4-5 Sims with usable AI ... That is good
 
Last edited:
I've also been looking into this AI behaviour and I'm starting to see the pattern, there are certain corners the AI consistently crashes on. Some tracks are fine others have a consistent problem such as Hohenrain on Nordschleife as shown in the video.

well, if they improve the AI-Behaviour a re-visit in the AI-Lines Department wouldn't be entirely wrong, maybe from someone who doesn't start braking inside the apex of a corner :D....the Silverstone GP part in that clip, where the AI is braking inside the apex, looks very familiar to me.
 
So why exactly are you here? Just curious why somebody that doesn't own the game, and that doesn't have any interest in the game, would waste their time commenting on that game?

As long as you are here though, which game do you find the ai interesting in?

...giving some feedback, voicing my opinion, a bit trolling, yeah, having a bit fun :laugh: and atm AC simply released a 'bigger update' with AI...as soon RaceRoom releaeses one, you'll probably find me there hanging around ;). (these are the 2 games i 'invested' back in the EA, Beta-phases).

GTR2 has so far the best AI, not perfect or ground-breaking, but solid......especially on original-tracks ,depeneding how the Modders created thei AIW's.

Out uf Curiosity: Does anyone knows, why the AI-Lines has to be driven first by a human which determines how fast or balanced the AI works on that track?! It's something i never understood in Racing-Sims, why you can't create a simple 'perfect AI-Spline' for a track and the AI-Physics simply drives along this 'perfect line'!?
 
Out uf Curiosity: Does anyone knows, why the AI-Lines has to be driven first by a human which determines how fast or balanced the AI works on that track?! It's something i never understood in Racing-Sims, why you can't create a simple 'perfect AI-Spline' for a track and the AI-Physics simply drives along this 'perfect line'!?
Because computers can't do anything automatically. If you want it to find the "perfect line" you have to write code that explains what the perfect line is, which at best is going to take hundreds of hours of work. It's a lot easier to just have a human record a good lap, which takes at most 20 minutes.
 
..giving some feedback, voicing my opinion, a bit trolling, yeah, having a bit fun :laugh: and atm AC simply released a 'bigger update' with AI
How is it possible that this guy is still able to write on this forum?
Out uf Curiosity: Does anyone knows, why the AI-Lines has to be driven first by a human which determines how fast or balanced the AI works on that track?! It's something i never understood in Racing-Sims, why you can't create a simple 'perfect AI-Spline' for a track and the AI-Physics simply drives along this 'perfect line'!?
So you understand almost nothing about computers and AI but you are always ready to talk **** about other people's work?
 
Out uf Curiosity: Does anyone knows, why the AI-Lines has to be driven first by a human which determines how fast or balanced the AI works on that track?! It's something i never understood in Racing-Sims, why you can't create a simple 'perfect AI-Spline' for a track and the AI-Physics simply drives along this 'perfect line'!?

If you're really interested in the subject and not against reading up & watching a few videos, Bink pretty much completed such a beast for rF2, but never released it:
https://forum.studio-397.com/index.php?threads/aiw-thing-video-6.47403/

*** Added ***
Also note that the original rFactor has an AI learning mode available to the player which optimizes the current AI path within the track boundaries. There are drawbacks to that method, in that it doesn't cater to the strengths/weaknesses of different cars and may take curbs too aggressively. This is called Type 1 learning in rFactor 2.

rFactor 2 builds upon this AI learning base with two more AI learning methods. Type 2 learning is available to the player and caters to the different cars by building Wisdom files (*.wis) so the AI doesn't make moves which destabilize the car. Type 3 learning is the newest and is primarily an improvement on Type 1 learning, but aimed at the developer so the user shouldn't need to use Type 1 learning.

Another thing worth noting about creating AI lines is that the width of the vehicle chosen to drive the lines matters. A wide car will not run optimum lines for a narrow car, but making the lines with a narrow car will typically put wide cars off the track edge. Similarly, creating AI lines with a low-grip car will leave too much unused track for high-grip cars.
 
Last edited:
Because computers can't do anything automatically. If you want it to find the "perfect line" you have to write code that explains what the perfect line is, which at best is going to take hundreds of hours of work. It's a lot easier to just have a human record a good lap, which takes at most 20 minutes.

i ment, you make the line for e.g. in photoshop or the track-tool, just the 'spline', so the AI allready knows the path to drive at, but just not recorded with any speed/brake points.... the AI itself determines the max. speed it can take the corner. It's like a overlay of a path over a track. The path doesn't contain any values, it's simply just the path and the ai simply follows the path with max. possible speed with it's available physics.
 

Latest News

How long have you been simracing

  • < 1 year

    Votes: 228 14.7%
  • < 2 years

    Votes: 157 10.1%
  • < 3 years

    Votes: 156 10.1%
  • < 4 years

    Votes: 117 7.5%
  • < 5 years

    Votes: 222 14.3%
  • < 10 years

    Votes: 183 11.8%
  • < 15 years

    Votes: 122 7.9%
  • < 20 years

    Votes: 84 5.4%
  • < 25 years

    Votes: 70 4.5%
  • Ok, I am a dinosaur

    Votes: 213 13.7%
Back
Top