Gran Turismo Sophy Goes Sideways

sophy_drift.jpg
During the latest official Gran Turismo 7 esports event in Amsterdam, a new demonstration of Sony's now famous racing AI was performed. It took everyone by surprise: Sophy was not just going super fast on an ideal racing line, it was drifting.

Image credit: Polyphony Digital

Some observers may have caught a glimpse of the demo before, but now, an official video is out. Sophy is producing rubber trails and smoke clouds on GT's iconic Trial Mountain track in a car you would definitely not expect to see in a Formula Drift competition - an old Lamborghini Countach.

Not content of doing it extremely well, managing to get great angle through a challenging high speed section, it also manages to look organic and take some artistic liberties by performing 360s in the middle of the run. More than the spectacular display of skill, it's the fact that what Sophy is doing in the video is not the optimal way of driving in any scenario in GT7. It is certainly not the fastest way around the track, but it's also not the best method to score points in drift mode either.


GT Sophy Drifts Explained​

The way to earn drift points is pretty straightforward in GT7: Basically, drivers want to combine angle, speed, and a good line, without going off track or touching a wall. But their section score will be null if they go above an angle of 90 degrees at any point, meaning the game will punish reverse entries or 360s to link two corners together.

Even in "free" drifting mode that excludes sections, you still need to properly settle the car to validate your current streak, and can lose those points if you 360 or let the back of the car overtake the front for just a fraction of a second.


No Background Info Yet​

What is really interesting in this regard is the AI's not bound by the regular rules of the game. It just does what it wants - or at least it looks like it. Surely, Sophy had to learn with incentives, albeit different ones than usual, but as interesting to know how this was done in deeper details, this part was ommitted in the video.

We are hoping for a more detailed explanation from Sony AI later on. Sophy managed to demonstrate its ability to tackle another complex task and impress pro players, so it did its job rather well. The fact that drifting is being explored in itself is also a good thing for the community, as drifting is often ignored in sim racing despite being a popular discipline, as numerous Assetto Corsa drifting communities can attest to.

What are you thoughts about this demo? Would you like to see more drifting content in more serious racing simulations? Let us know in the comments below!
About author
GT-Alex
Global motorsports enjoyer, long time simracer, Gran Turismo veteran, I've been driving alongside top drivers since the dawn of online pro leagues on Gran Turismo, and qualified for the only cancelled FIA GTC World Tour. I've left aside competitive driving in 2020 to dedicate myself to IGTL, a simracing organisation hosting high quality events for pro racers and customers, to create with friends the kind of events we wished we could have had. We strive to provide the best events for drivers and the best content for viewers, and want to help the simracing scene grow and shine further in the global esports scene.

Comments

Like Richard said, getting a computer-controlled car to drift convincingly is impressive!

Ultimately, I hope Sophy moves beyond its current status as a research project. It would be amazing to see it integrated into GT7... or GT8. The racecraft abilities of Sophy AI (when not driving sideways) would make for immersive single player racing.

I still have two concerns with Sophy, though:

1) Can it remain convincing when its ability is "scaled down" to match the skill level of the player (if this is possible at all)? Be it a Sophy-controlled opponent in a drift mode, or Sophy-controlled AI cars in a circuit race, to my knowledge, we've only seen Sophy AI performing at its highest possible level. But if Sophy is to be used for all AI racing in a GT game, difficulty scaling is required. One disadvantage of the sophisticated machine learning Sony is using – it is challenging to tweak parameters and input data in such a way to predictably change the resulting outputs. It's not impossible that attempts to slow down the AI would lead to unusably bizarre AI driving behaviour.

2) Can Sony's sophisticated machine learning "scale up" to all scenarios and car-track combinations? I hope so, but I fear perhaps not. So far, it appears that each car-track-scenario combination has required a very large amount of time (and computational power) spent re-training models. Depending on precisely how their machine learning training works "under the hood", this could preclude Sony/Polyphony from being able to use Sophy AI across an entire GT game.
 
Last edited:
Good for you, how is this relevant to the article?

AC's AI can't even go straight properly, much less sideways...

Its a big achievement AI wise, i really hope Sophy becomes the de facto GT AI soon.
As much as I love AC. That game's Ai is literally non-existent. Would love to see a sim with real AI on PC soon.
 
It is a interesting concept from the physics, and coding point of view. And totally boring to watch a computer that has instant reaction times, overwhelming computing power over an human, and can even calculate what the car is going to do with exact precision just moments after the simulation started with the car still stopped on the track.

To me it is as boring as those Beamng videos in where someone is spinning and drifting in weird ways surrounded of cones extremely close to the car at all times. Then you discover that all is fake because the driver just spins a car and drift it on a parking lot and after he has the replay he edits it to add the cones.

I want to see a IA driving with the human limitations in input speed, input lag, reaction times and human levels of precision. Once you teach an AI to drive as the developer want's it to drive it is a task as easy as to us to breath.

But, that said, I believe that poliphony should further upgrade it and license it to F1 teams to test all the possible setup combinations for any track in mere minutes or hours. Any F1 team offered something like this would pay quite a lot for it.
 
I want to see a IA driving with the human limitations in input speed, input lag, reaction times and human levels of precision. Once you teach an AI to drive as the developer want's it to drive it is a task as easy as to us to breath.
I fully agree however it's sooooooooooo much more than reaction times / input speed. The AI has near-perfect skill. It knows almost exactly how to manipulate the car's brakes, throttle, steering in order to get the car to behave almost exactly how it needs to in order to get around a track. I personally don't have the quickest reaction times, and I know some people who have very quick reaction times (we used to do reaction test games a lot with digital watches, computer games, etc.) yet I destroy them in sim racing. You can slow down "Sophie's" reaction times by a lot but her driving will still be fantastic because she has incredible car "skill" and "feel" along with unbelievable amounts of "knowledge" of exactly what the tyres are doing, what the chassis is doing - the entire car and every single aspect of the car for that matter -, what the corner/straight/track requires, etc. She can receive and comprehend the car's full telemetry - every single trillion of trillions of bits of information thing that is modelled in the entire physics engine - and she can understand every single trilion of bits of information she is receiving.
 
Last edited:
1) Can it remain convincing when its ability is "scaled down" to match the skill level of the player (if this is possible at all)?

They are doing that already. Final image state, unrestricted, is about 2 - 3 seconds a lap faster than the best GTWS drivers. They had a race where its pace was more closely matched. Then the public version we got in the game for the demo was maybe a tad slower, or at least similar - close to the best laps of the fastest human players.

There's at least a couple ways to do that I can think of:
- take an earlier image (the neural network learns by repetition and generating a new version of itself each time)
- adjust some of the reward / punishment conditions

I want to see a IA driving with the human limitations in input speed, input lag, reaction times and human levels of precision. Once you teach an AI to drive as the developer want's it to drive it is a task as easy as to us to breath.

Sophy is already limited in input frequency, and literally operates a controller. Now regarding precision, an AI is a program so it will just execute what it's coded to do perfectly no matter what, so you can't get prevent it from being as precise as its environment allows it to be. If you want it to appear less consistent, you have to somehow create the illusion by forcing it to do so. Could be litteral RNG to trigger an event, could be a penalty for repeating the exact same thing twice, or a lot of different creative ways that would probably add even more complexity to it.

They are definitely going in the right direction though. You need the AI to be capable of the best if you want other scenarios to also feel real. And there's nothing Sophy does that a human technically can't replicate. Some of the fastest guys did adjust their driving style based on Sophy's and managed to replicate some of its craziest moves.

Remember Sophy is made by Sony AI, not Polyphony themselves. It's a reasearch project first, Sophy being a future feature for Gran Turismo is a byproduct of that.
 
I laughed when I saw it spinning 360's. I'm sorry, I find it impossible to be impressed by that. It looks like the kind of clown show stuff you'd find in The Crew or Forza Horizon. I get that under the hood it's probably a fantastic achievement for AI, but I hope they actually put it to good use someday.
 
I laughed when I saw it spinning 360's. I'm sorry, I find it impossible to be impressed by that. It looks like the kind of clown show stuff you'd find in The Crew or Forza Horizon. I get that under the hood it's probably a fantastic achievement for AI, but I hope they actually put it to good use someday.
What is so funny in 360 drift?
 
They are doing that already. Final image state, unrestricted, is about 2 - 3 seconds a lap faster than the best GTWS drivers. They had a race where its pace was more closely matched. Then the public version we got in the game for the demo was maybe a tad slower, or at least similar - close to the best laps of the fastest human players.

There's at least a couple ways to do that I can think of:
- take an earlier image (the neural network learns by repetition and generating a new version of itself each time)
- adjust some of the reward / punishment conditions
Thanks for the information! :thumbsup:

An earlier image is an interesting idea. Perhaps it could work. The risk probably being lower quality driving and racing behaviour (if the network hadn't 'learned' enough yet). Just a guess, but perhaps your second idea, adjusting rewards and punishments, is better. Perhaps by including a difficulty value as an input, training on data from slower human drivers, and punishing if the computer-controlled car is too fast or slow compared to a target laptime (which is congruent with the difficulty value).

Ideally, though, you'd be able to continuously adjust the AI speed on a percentage slider to match your pace. Would be interesting to know if that would be possible, given how Sophy's training works, or whether this would be too intensive. if nothing else, I bet they could train a few different difficulty levels: say, "Easy", "Medium", and "Hard" modes. :)
 
They have talked about difficulty levels, and their language hinted more towards using images at different stages of learning to reflect not only the pace, but also the behavior of less proficient drivers. I have the feeling it may be the biggest challenge to come.
 
Ehm...what is the news here? Just program a car to drift? Any dev can program this. Noboby is telling me this a computer ai which is drifting because its fun....
 
Ehm...what is the news here? Just program a car to drift? Any dev can program this. Noboby is telling me this a computer ai which is drifting because its fun....
The AI learns to master drifting with the car themself, otherwise you should program new drift values for every track, every car and it's countless upgrades.
 

Latest News

Article information

Author
Alexandre Tonini
Article read time
2 min read
Views
3,349
Comments
15
Last update

How are you going to watch 24 hours of Le Mans

  • On national tv

    Votes: 253 34.1%
  • Eurosport app/website

    Votes: 202 27.2%
  • WEC app/website

    Votes: 140 18.8%
  • Watch party

    Votes: 61 8.2%
  • At a friends house

    Votes: 18 2.4%
  • At Le Mans

    Votes: 69 9.3%
Back
Top