Gran Turismo Sophy - A breakthrough in AI?

GT Sophy.jpg
Gran Turismo Sophy is a new take on AI within a driving game. The project is a collaboration between Sony AI, Polyphony Digital, and Sony Interactive.

According to the development team, GT Sophy is the next level to AI and recent results would suggest they are right. They are claiming GT Sophy is an AI breakthrough, comparing this AI to other breakthrough AI that have mastered arcade games, chess, and multiplayer strategy games.

Does this mean that dive bombing AI into T1 and making up half a dozen places in a thing of the past? Or does this mean bullying AI out of the way is a tactic we’ll no longer be able to do?

The GT Sophy project has been in the works for six years, where GT Sophy was trained using deep reinforcement learning techniques. The development team didn’t build an AI to cheat the game, like many other AI do - they developed an AI that learns by completing human-like tasks. The reinforcement techniques “reward” or “penalize” the AI for specific actions it takes inside an environment.

GT Sophy's learning covered three areas of driving skills:​


Race Car Control - Operating at the edge
GT Sophy acquired a deep understanding of car dynamics, racing lines, and precision manoeuvres to conquer challenging tracks.

Racing tactics - Split-second decision making
GT Sophy showed mastery of tactics including slipstream passing, crossover passes and even some defensive manoeuvres such as blocking.

Racing Etiquette - Essential for fair play
GT Sophy learned to conform to highly refined, yet imprecisely specified, racing etiquette rules including avoiding at-fault collisions and respecting opponent driving lines.

On two occasions last year GT Sophy competed against some of the best Gran Turismo drivers

In July 2021, GT Sophy had the fastest lap in all three races and won two of the three races. Yet the human GT drivers won with the most points.

During October 2021, GT Sophy won all three races and finished second in all races.

GT Sophy has achieved a major milestone however PDI, SIE, and Sony AI will continue to develop and upgrade their AI’s capabilities as well as explore how this AI can be implemented in the Gran Turismo series in the future.

Additionally, Sony AI is also exploring new partnerships to enhance the gaming experience for players through AI.

What are your thoughts about GT Sophy? Is it about time AI in racing games took the next step?
About author
Damian Reed
PC geek, gamer, content creator, and passionate sim racer.
I live life a 1/4 mile at a time, it takes me ages to get anywhere!

Comments

Is the Sophy AI actually in GT7? I was under the impression it’s a separate project using GT7 as the platform for the research.
 
If we see less "monster grip Ai drives on rails" i definitely welcome it!
I think about scenarios where you have to lift the gas to maintain grip, but the Ai is just pushing flat out thru corners... if this is gone, nice!

Or the opposite:
On some tracks, the Ai is breaking every round exactly in the same spot and depending on difficulty too early, what results in a "cheap" racing experience.

There are a lot of factors where this could improve racing quality.
I am definitely curious how this will evolve and might adapted/used by others. :thumbsup:
 
Last edited:
As I understand it then the other neural network based AI MotoGPs ANNA isnt in practice working enourmously ;) different from todays rather lousy Artificial un-Intelligent bots.
But if they can get it to function (ANNA and Sophy & comming friends) then it maybe for the first time could be fun and challenging to compete with some AIs in racing games that feels Intelligent.:roflmao:
 
Last edited:
I think this will be a gamechanger for racing against AI. Not only will racing against this type of AI be more realistic (the top GT players thought they were racing against real people), but this type of AI could also be used as a learning tool in simulators. Sophy not only defeated several world class GT players, but it also showed HOW to drive faster around a track. At the Dragon trail track the Sophy AI rode a curb in a braking zone in order to take the next turn faster. Very interesting to see. I hope this type of AI also becomes available for other racing game developers.
 
I take this with a massive pinch of salt, given the GT series AI has been carp for 20 years. You can put any gmotor ISI based AI up to 120% and it will smoke the fastest human players around too.
When a developer can make AI as good as GP4's as a starting point, them I might believe all these fancy 'neural' PR announcements. As it is, no title has got close to what Microprose managed 20 years ago, yet alone made it as believable as a human competitor...
 
I take this with a massive pinch of salt, given the GT series AI has been carp for 20 years. You can put any gmotor ISI based AI up to 120% and it will smoke the fastest human players around too.
When a developer can make AI as good as GP4's as a starting point, them I might believe all these fancy 'neural' PR announcements. As it is, no title has got close to what Microprose managed 20 years ago, yet alone made it as believable as a human competitor...
No need to take it with a massive pinch of salt. Sophy does already exist, drives according to the racing etiquette, is already better than the world class GT players and most importantly: races with the same physics as the player. There are several videos online where you can see Sophy in acual races against top drivers. Only problem it is not implemented in GT7 yet is that they haven’t figured out yet how to make it slower, so the average joe is able to beat it.
 
Last edited:
I love that GT is pushing for innovation in AI racing and trying new things. About time someone did this!

I am a bit concerned, though, about how Sophy's abilities will 'scale' to various difficulty levels lower than world-beater. Because I use AI in lieu of racing humans, I need AI that are as slow as I am :laugh: and whose speed can be scaled in a relatively convincing way – something Papyrus and ISImotor AI, in general, does quite well. Not some bonkers neural network that has learned to drive to almost perfection and uses some crazy-ass line through the Ford chicanes at Le Mans cause the devs didn't specify the track boundaries correctly.

When a developer can make AI as good as GP4's as a starting point, them I might believe all these fancy 'neural' PR announcements. As it is, no title has got close to what Microprose managed 20 years ago, yet alone made it as believable as a human competitor...
While in my experience the GP4 AI doesn't stand up quite as well as nostalgia would tell us, it is very sophisticated in terms of its racecraft and pit strategy. Why? Because Geoff Crammond explicitly hard-coded it the "old-fashioned way" to behave like this, not some statistical learning model working things out for itself. I can see using machine learning models to allow the AI to choose correct inputs to stay on a given path... but that's it; anything more seems to be overcomplicating things unnecessarily and introducing the potential for poor or un-lifelike results. In my view, things like pit strategy, situational awareness while racing door-to-door, different racing grooves need to be manually specified based on heuristics used by humans to be convincing to human players driving against AI.

EDIT: I don't mean to be overly "Bah humbug" about Sophy though... I'd love to be proven wrong!
 
Last edited:
Fortunately, GT is a simcade cause i'm really curiouse about how such an agressive driving and cornering could handle tire wear in a long race in a real simmulation like RF2 or ACC ! All that sophy's tyre will be totaly destroyed in a few lap .....
 
Fortunately, GT is a simcade cause i'm really curiouse about how such an agressive driving and cornering could handle tire wear in a long race in a real simmulation like RF2 or ACC ! All that sophy's tyre will be totaly destroyed in a few lap .....
Probably would handle it quite well, since i have still to come accross any "sim" that actually penalizes agressive fast driving (this is different from making mistakes while driving) with more tire wear. No, not even those two pull it off convincingly. RF2 tire model has the potential for it, but i haven't seen a proper implementation.


EDIT: sophy does seem very impressive from the videos i've seen. Its good to see real effort coming into the AI realm.
 
Last edited:
Fortunately, GT is a simcade cause i'm really curiouse about how such an agressive driving and cornering could handle tire wear in a long race in a real simmulation like RF2 or ACC ! All that sophy's tyre will be totaly destroyed in a few lap .....

I was thinking the same while watching the races.

Another thing that comes to my mind is "reflexes". It's great Sophy learn and drive by using the same physics as the human player, but if her inputs are at AI blazing speed, there is no point to compare against humans in my view.

As others had mentioned, learn to drive is one thing, but to be complete she needs to learn tyre management, pit strategies, etc.

Anyway, it's very good to see some real advancements in the simracing genre regarding AI. Lots of advancements everywere, but AI in current titles are even worse than 20 years old games.
 
We don't need super human AI that beats the best players. We need AI that feels human and make mistakes and not 100% always flawless. It also has to accommodate to average drivers for providing them a challenge...
Also, do we know if Sophy is actually using game physics (vs just being a dumb bot going at speed X, direction Y if track conditions are A,B,C,D, ... using just pre-computed data during its training) ? And how do we know if AI is not cheating, doing something that no human can realistically do (assuming it stays within the physics of the game) ?
 
Last edited:

Latest News

Article information

Author
Damian Reed
Article read time
2 min read
Views
10,362
Comments
54
Last update

Do you prefer licensed hardware?

  • Yes for me it is vital

  • Yes, but only if it's a manufacturer I like

  • Yes, but only if the price is right

  • No, a generic wheel is fine

  • No, I would be ok with a replica


Results are only viewable after voting.
Back
Top