UZH Robotics and Perception Group
Autonomous car racing raises fundamental robotics challenges such as planning minimum-time trajectories under uncertain dynamics and controlling the car at its friction limits. In this project, we consider the task of autonomous car racing in the top-selling car racing game Gran Turismo Sport. Gran Turismo Sport is known for its detailed physics simulation of various cars and tracks. Our approach makes use of maximum-entropy deep reinforcement learning and a new reward design to train a sensorimotor policy to complete a given race track as fast as possible. We evaluate our approach in three different time trial settings with different cars and tracks. Our results show that the obtained controllers not only beat the built-in non-player character of Gran Turismo Sport but also outperform the fastest known times in a dataset of personal best lap times of over 50,000 human drivers.
Reference:
F. Fuchs, Y. Song, E. Kaufmann, D. Scaramuzza, P. Duerr
Super-Human Performance in Gran Turismo Sport Using Deep Reinforcement Learning
IEEE Robotics and Automation Letters (RA-L), 2021
PDF: http://rpg.ifi.uzh.ch/docs/RAL21_Fuchs.pdf
More about our research on Deep Learning: http://rpg.ifi.uzh.ch/research_learning.html
Affiliations:
Y. Song, E. Kaufmann and D. Scaramuzza are with the Robotics and Perception Group, Dep. of Informatics, University of Zurich, and Dep. of Neuroinformatics, University of Zurich and ETH Zurich, Switzerland
http://rpg.ifi.uzh.ch/
F. Fuchs and P. Dürr are with Sony, Switzerland.
Source
Holy .you can apply it to anything
0:12 and noticed clipping. xD
Try getting the same results taking in pixels as inputs.
Now take this and apply it to the penalty system and improve it because it needs work. It's no bad if you consider all the variables and how hard is to program so many possibilities but now it's time to go super human and put AI to the task…
Curious what the input data was. I'd like it to be just visual, and with a defined reasonable resolution similar to a human's resolution. It would be interesting to give it a defined resolution of its outputs too, especially steering.
And then give it force feedback inputs and compare.
Although this is impressive, I wish to point out the AI have an inherent advantage here because its access to the game's "Feature Extraction" effectively gives it a live 360* surround sensing, like doppler parking sensors, which is what truly allows it to get exactly to the edge of the walls. If human players had something similar, a small parking sensor type display, they too could in time match this level of performance. But, human players do not have this advantage. As another comment suggested, let's see this AI again but with an input sensing methodology closer to what a human relies upon (and is limited to – touch, vision, audio). Bravo anyway, and thank you for sharing this video.
GREAAAAAT now the Super license will be even harder to get on the new game. Cheers SkyNet for being self aware
Would be interesting to know what kind of parameters both human and AI used. Would also be interested to see this AI try to run a race simulation with fuel and tyre wear. Now what would be even more interesting is having the AI run a race with dynamic conditions in another game, and have it run on Assetto Corsa as well, which has incredible physics, and a more in depth tyre model.
WE WANT MORE!!! A playlist with more driving tracks and with inside and outside views… 😃😃😃
Nice.
This only works with stable settings (no Tyre wear, no fuel consumption) i guess?
End human driver end the planet real
Can you show us the inside cockpit. I would like to know steering angle.
Just wondering, does the AI control gran turismo directly or is it using a controller like a human. If its the former then it's very good but superhuman?
When self driving cars become standard we'll wonder how many lives have been saved by not continuing to drive cars manually….one day
can you guys make a mod for pCARS2? The AI needs some love there as well..
I think the lap is beatable though.
Less than 0.2 second gap.
It'll be interesting challenge for sure.
No we will know what is the best car for each track
Is this who is going to take the championship from Lewis Hamilton?
I wish they put that efforts in their consoles
sport.ai races.ai
The AI didn't discover anything new, it just took advantage of computer's precision and speed that a human can never match.
OK, next project: New PUBG AI wins the chicken dinner against pros
Imagine if regular, consumer friendly games get AI like this for training its players for online without going online.
Not only it would create fun, but also posing challenges that can give satisfactory results when beaten. However, cheating can still be exploited here.
Tell the ai to do a regular daily race.
First person to beat this AI is gonna get taken away for 'tests' and never seen again.
let the ki try iracing against an autistic fin that uses a 1k $ wheel and VR
？？TAS？？？
Now if only the GT7 AI is like 70% of this, I'll be a happy man. A blind person could win against the level of GTS AI.
Interesting. Would like to see your future work where multiple cars are in the track & your RL model outperforms others human drivers present in the track. Best wishes.
How is it performing in a real race with other opponents?
Please… For the love of God… Let this be a proof of concept.. And not the ai in the next game..
Interesting… but when GT7 will be released, KI will be as stupid as always 😀
Obviously the agent knows the curb is fake!!!
can i borrow this to use in forza horizon 4?
there's this prize i want to get but i'm having trouble improving from 250th place 🙁
@UZH Robotics and Perception Group
Driver's choices are infinite so how do you model action state spaces?
why can't polyphony do a good AI?
an elementary task, what is it proud of here…?
要は理論値を出すためのAIか
The problem with this is when you change the track, the AI just won't work since the model is probably overfitted…
That's another story if there are other opponents
This is interesting ! How did the feature extraction works ? Polyphony Digital have opened their API so you can get data ?
This is amazing, I presume the API used was for private use only ? Meaning we cannot reproduce the experiment ?
we also did this in our group thesis way back in 2018.our approach is we use projector and camera tracking as a reference to the input of the controller.
94947 on Nov 21 2021 NOON Chicago time zone
The perfect lap finally arrived
"Unplug it, fast!"
This was 1 year ago and PD already announced Sophie