For the first time, a single machine learning model has begun to learn how to drive two totally different types of vehicles: a passenger car and a delivery van. It was the creator of this artificial intelligence Wayve, a London startup.
Because it is important
Less than a year ago, Wayve demonstrated that he could use artificial intelligence 'training' on the streets of London, and use it to drive cars in four other British cities as well. A short time ago this goal would have required a lot of time and effort.
And it comes close to human behavior: after all, learning to drive is independent of the place. You drive in your city, then you go to another city (maybe you hire a car) and you can drive anyway, can't you? I made it simple, but in essence the news is that this technology is far, far ahead of current players like Tesla, Waymo and others.
A new generation that knows how to learn to drive
The difference between Wayve and the others is in the approach. It is a much smaller company than its better-funded competitors. But it is part of a new generation of startups, which some call "AV2.0", which is moving away from the robotic mindset embraced by the first wave of companies developing autonomous vehicles.
Until recently, the approach was based on super-detailed 3D maps and separate modules for surveying and planning. Waywe relies entirely on artificial intelligence to drive vehicles. In other words, the car has its own criteria for learning to drive, and it adapts without having to resort to maps. Learn, interpret, practice.
Train and drive several different vehicles, even at the same time
The initial approach to autonomous vehicles he carried around a few prototypes, with great effort and cost. I don't know if it will spread in these terms. Wayve and other next-generation startups want to repeat with cars what deep learning has done for natural language processing.
Wayve vehicles are equipped with the same sensors as similar cars, but positioned higher and at different angles. And the first difference stands out: with this approach, it is the artificial intelligence that manages everything. If I place these sensors on a small car, the AI calculates the difference in angle and "learns" to handle the different dimensions of a vehicle. Do I put these sensors in a truck? The AI adapts to the truck. And it adapts to everything: size, mass, braking time, steering angle.
In other words: how a human being changes modes depending on the type of vehicle he drives. He "understands" that he is in another vehicle, and can learn to drive accordingly. This has several consequences.
How this model of artificial intelligence works
The AI model of Wayve, Ghost, and other next-generation startups is trained using a combination of reinforcement learning (where it learns by trial and error) and imitation learning (where it copies the actions of human drivers). It took thousands of hours of driving data to train the model to drive a car. Soon after, the artificial intelligence was "put to drive" a van. Result? Only 80 hours. And it has also improved in driving the car before.
"I felt a little scared during the first ride of the van," she admits Naomi Standard, the operator Wayve who was sitting in the driver's seat, not driving. "I felt like a driving school instructor with a novice student." The van, however, coped well with London's narrow streets, proceeding correctly through road works, walkways, double-parked cars and other obstacles.
Imagine now what will happen soon. These AIs will be placed on 20, 30 different types of vehicles, "re-learning" to drive on each of them in a slightly different way. And by putting these modalities together, derive a "general model" of driving.
The future: learning to drive
Ghost wants to make consumer vehicles that can drive alone on motorways; Wayve wants to be the first company to put driverless cars in 100 cities.
Both will contribute to a possible integrated future. An artificial intelligence service to which our autonomous vehicle will already be registered (a bit like a smartphone is equipped with an android system): an artificial intelligence that will drive many vehicles at the same time, adjusting their behavior in order to guarantee perfect safety.
Advanced artificial intelligence will create a standard, a single "brain" that will have control of everything that drives on the road: let's expect a long time for this, of course. Not for technical reasons, though. For ethical and human reasons. A similar system, in fact, presupposes the existence of roads where no "human" vehicle can circulate. A not negligible detail.
At least for now.