Gran Turismo: How deep learning will revolutionise AI in sim racing games


There's not that long to go now, Gran Turismo 7 will soon be upon us. Polyphony's first foray onto the PlayStation 5 will launch in March next year and we can't wait!

This is thanks in part to the fact that Polyphony is always pushing the envelope when it comes to technology. Even on the dawn of a new game like we are now, the developers are looking into ways to improve.

A recent experiment shows we could be on the cusp of a revolution when it comes to programming AI. What implications could this have for GT7 and racing games though? We've got everything you need to know right here!

The previous study

Gran Turismo Sport tokyo expressway
expand image
CLOSE TO HOME: The Tokyo Expressway was chosen as the circuit for the tests to be carried out

The study in question was carried out Sony AI - a new subsidiary of Sony that studies the applications of AI. This was a follow-up to a similar study from August 2021, but with a few small, but key, differences.

The new paper posted to the arXiv research repository demonstrates the possibility of teaching AI how to race in GT. This is by using the same visual cues as human drivers.

In the previous study, reinforcement learning was used to teach AI to go as fast as it could be based purely on physical interactions with the environment. The AI learned which paths resulted in collisions and slower lap times, and eventually achieved “super-human” performance.

Second attempt

Gran turismo 6 snow
expand image
IF AT FIRST YOU DON'T SUCCEED: Try and try again...

The new paper focuses on visual cues instead. Ryuji Imamura, the paper's main author, hypothesised that in theory, the information conveyed while driving — the track itself, the head-up display, and force feedback — are sufficient for a human driver to learn how to drive quickly. Therefore, it shouldn't be any different for an AI to learn to drive at human speeds.

The experiment ultimately consisted of a process of running an AI driver from an initial lap through several hundred batches (or “epochs”) of laps. The AI driver observing its laps through a capture device and from speed and acceleration data provided through the PS4 console itself.

After each set of tests, the new data would be analysed and used to generate new control outputs for the next set.

The lap time data for each epoch was compared to a batch of 28,000 user-generated lap times held by Polyphony Digital. The event was an arcade mode time trial at Tokyo Expressway Central Outer Loop in the Mazda Demio XD Touring ’15.

The results

Gran Turismo Sport Interlagos 01
expand image
RAPID: The AI in this experiment were very quick

The visual learning AI quickly beat the lap times generated by the GT Sport AI, getting to that mark in around 20-30 epochs — equivalent to 350-500 laps. It didn’t take much longer to exceed the median lap time of the human players either. Although, improvements stagnated at around 200 epochs, or about 3,500 laps.

Ultimately, by using more-or-less the same cues as a human player, the visual learning AI managed to achieve a lap time within the top 10% of players. This was around 3.3 seconds slower than the fastest human, 4.5s faster than the median human, and a whopping 9.5s faster than GT Sport‘s own AI.

What this could mean for the future of racing

gran turismo laguna seca
expand image
BEST DAYS BEHIND THEM: GT's AI hasn't been up to scratch for years now

A lot has been made of GT's AI in recent times. There's no getting away from it, it hasn't been the best for a long time. This isn't without good reason though, as Polyphony have focused more on the tracks and cars than the AI.

For GT7, Polyphony has reassured players that the AI will be better than in previous games. However, we don't know exactly how this will be true. Furthermore, until we actually play the game next year, we can't say whether that will be the case.

This study though, has far-reaching implications and could be revolutionary in racing gaming. Creating an authentic AI in racing games isn't as simple as setting a lap time for them to adhere to and letting them get on with it. If that was the case, every game would be perfect.

It's not just how fast they are, if that was the case then the first study would've been a success. The AI also need to be fast in the right places, otherwise the pace difference between the players and them would be too large. By using the same cues as real players to improve its driving style, the AI improves in the same ways players do. So, the gap between AI and human players will all but disappear in time.

This won't be something that will impact GT7, but going forward, it can and should be a game-changer. Watch out for developments on this ahead of GT8, whenever that will launch.

This Article's Topics

Explore new topics and discover content that's right for you!

Gran Turismo