AI is the talk of the town right now, with the likes of Google Gemini scouring the internet for answers or Morgan Freeman-aping voice generators. If you’re on the hunt for a slightly weird-looking man riding a bicycle in 2056, there might be an AI-powered tool to help with that task.
But that seems a little much, like putting ancient Rome expert Professor Mary Beard in the Sunset Selling event.
Artificial intelligence can be used for more than making random wedding speeches – we may not have seen the best of it yet.
In competitive video games and e-sports, players have long been familiar with ‘AI’ – for us, the basic concept is nothing new. We’ve been racing on the ‘computer’ for decades, with the game and the hardware that runs it making your rivals go around the same track as you.
Some, no doubt, are better than others. A direct criticism of the game or the simulation is ‘the AI ​​is bad’, with a lack of awareness, hitting the first lap or repeatedly crashing at the same point.
That’s because, at a basic level, never really learn. They cannot understand wChickens make mistakes and they struggle to change their behavior based on what is happening around them.
When Gran Turismo Sophy launched Gran Turismo 7 to the public in late 2023, the secret agent was a revelation. Opponents can attack, defend and avoid potential events more naturally.
The key difference between the pre-programmed AI that many have used, and an agent like GT Sophy, is its ability to adapt to the context.
It’s been the culmination of almost five years of work, so far, with the July 2024 update for Gran Turismo 7 expanding the list of tracks it can race on.
Sony’s new AI racing agent with eyes
But, for Sony’s global AI team, that’s not enough. It has been working diligently as an exploratory AI agent, just like GT Sophy.
Again with Gran Turismo being used as a test platform, still in prototype stage, the way it receives and interprets information is different – it can ‘see’.
“When you or I play a game, we’re, of course, looking at the user interface and parsing that information into ways that make sense to us, and we use that input to drive,” said Kaushik Subramanian, Senior Research Scientist at Sony. AI.
He collaborates with Takuma Seno, a Research Scientist and one of the Lead Authors of a recently published research paper on the latest developments.
“However, when we train GT Sophy, we give it very accurate information about where it is on the track; its running pattern, its exact speed, its acceleration and where other cars are.
“So there’s a difference between what Sophy takes and what human players take.
“What we wanted to explore is if we could train an agent to work with the imaging tools.”
“If I take a picture of what the game looks like – a car driving on a road with track lines, obstacles, trees, and clouds – and try to train the racing agent with that idea, is that possible?”
So that’s exactly what a survey agent uses. Imagine a camera sitting in a car or two eyes (an ego-centric camera view in Sony AI parlance). The new AI ‘looks’ at the car and interprets what it ‘sees’, experiencing the speed in addition to the data in the form of the correct speed.
Therefore, the goal is to match what people can find while playing the game, with the image helping to explain the style of the track. Unlike GT Sophy, it is not fed with information such as exact track shape, tire slip angle (and slip ratio) or tire load.
“One of our internal motivations was to try to see how much we could relax some of the information entering the agent,” continued the PhD graduate of the Georgia Institute of Technology, who has worked on Sony AI since 2019 on Gran Turismo. -next projects outside.
“The interesting part about this is that we don’t control what it focuses on. We provide the image, and the AI ​​process learns what to focus on to get the result we’re interested in.
“We tell the AI ​​that its goal is to drive as fast as possible, stay on track and avoid crashing into obstacles.
“Given some training time, the algorithm learns to focus on the relevant parts of the image to satisfy that goal. In the direct part, the AI ​​focuses on what is coming out of the distance.
“Whereas if it is looking from an angle, it is marked with red dots [used to determine its focus] he may line the edges of the track to help her stay within the bounds of the course.”
The world’s premier educational agent for integrated car racing
Currently, the new model is only working in a time trial situation, although “It seems to be a logical next step to now take this work and improve it in other situations, including races against cars,” Subramanian theories.
A fast swinging AI on the track is one good thing, helping it understand the ethics of racing is another.
As it stands, this is one car, one road, and for example, it has learned to go much faster than most human players. It can take the perfect line, millimeter perfect, every time.
This follows an exercise lasting 10-14 days on a group of PlayStations with access to Gran Turismo 7 – this new model is by no means guaranteed to make it to retail versions and is still in an exploratory phase.
During this time, the AI ​​uses reinforcement learning to improve. This, in essence, is a form of trial and error, starting out casually but eventually becoming a master.
“In the end, the result is what we refer to as a policy, a policy that works as a control,” said Subramanian. While in Germany, he co-authored two research papers on racing and one on driving.
“After two weeks of training, the supervisor is ready and has learned how to drive the best lap. I press start. It just gives me the lap time on the first attempt. “
The net result, so far, is an agent that can deliver roughly the same times as the existing GT Sophy but without the need for external hardware. This is the creation of the first human agent that represented the car race with local sensor input.
You can’t help but get the feeling, however, that the Sony AI team would like to push the boundaries even further.
“On our side, there are several goals, including several internal research questions, that we are trying to push to the limit,” said the experienced enforcement expert.
“Expanding the perspective of other situations is definitely interesting for us. We will try to do as much as we can in that position.
“Obviously, there’s GT Sophy deployed in the game as well, and that’s an interesting thought in our minds to make sure players are happy with that version of the game.
“When we started the project, we weren’t sure if it could actually be learned, because we know humans can do it, but the question was, can AI do it too? It took a while, a few years to get to where we were. where we are now, and the results so far have been promising.”
Sony AI’s latest autonomous racing agent in this regard
- The Sony AI team used PlayStations and Gran Turismo 7 to evaluate the exploration agent, but it has not been confirmed for public release.
- A different agent AI Gran Turismo Sophy, currently in the research phase
- The first human-based super car racing agent
- It can outrun the best Gran Turismo drivers on the planet in one lap
- Unlike GT Sophy – which relies on external inputs such as the exact shape information, slip angle, slip ratio and tire load (among others) – the new agent can ‘see’ the pixels (‘image’) through the ‘ego- centric view of the camera’ along with the availability of speed and acceleration information.
#Sony #created #human #agent #racing #Gran #Turismo