Liu’s team uses machine learning and real-world data to dramatically reduce the time it takes for self-driving cars to become seasoned drivers in complex road conditions.
Image: Brenda Ahearn, University of Michigan
Based on the recent hype about autonomous vehicles (AVs), our roads are traversed by hordes of driverless vehicles, whisking vehicle owners to their next destination, where they spread real and fake news. You would imagine trying to tell them apart. Cell phone in the back seat.
But convoys of self-driving trucks, self-driving taxis and fleets of passenger cars have not materialized like the breathtaking hype of a few years ago.
Also, self-driving cars that greet you when you wake up may be your future
The Insurance Institute for Highway Safety expects just 3.5 million self-driving cars to be on U.S. roads by 2025. That number will reach up to 4.5 million by 2030. Instead, it relies on human counterparts to make key decisions.
So what’s holding back self-driving cars?
not easy
Believe it or not, even though these self-driving cars have driven millions of miles for testing purposes and are equipped with all sensors and highly detailed city mapping, self-driving cars are I can’t seem to run it. The only important thing they should do is anticipate the glorious unpredictability of other volatile humans.
Henry Liu, a professor of civil engineering at the University of Michigan, told ZDNET, “The safety performance of these self-driving vehicles is currently not on par with human drivers, even with state-of-the-art self-driving systems.” I’m talking .
again: This new technology could blow GPT-4 and others.
Liu is also the director of M-City, a 32-acre mock city on campus that tests self-driving cars, and the director of the Connected Automobile Center funded by the U.S. Department of Transportation.
As Liu explains, the main problem with AV hogs is the “curse of scarcity.” It is a fact that we rarely encounter accidents in everyday driving. Self-driving cars would need to drive hundreds of millions, possibly billions, of miles to encounter and learn from some accidents.
For example, self-driving car company Waymo has reached 1 million miles of public self-driving without any human monitors in its vehicles, with 18 minor “contact events” and two major “contact events.” Consider the case where you are advertising the fact that you have only experienced it.
GM also unveiled a semi-autonomous driving system priced at $300,000.
How will this car react to someone who decides to cross the road on a whim, like a child running late on the morning trek to school?
An incident of this kind actually happened in Tempe, Arizona in 2018. Uber’s test vehicle could not identify a person crossing the road on a bicycle outside of a pedestrian crossing, did not take evasive action on her, even if it was possible, and crashed into her and died. I was allowed to.
Today, self-driving cars have constant problems identifying objects on the road, from paper bags to flocks of pigeons. Otherwise the results are fatal.
Self-driver edition
The highway portion of M-City’s augmented reality test environment for self-driving cars.
Image: Brenda Ahearn, University of Michigan
Humans must negotiate random, complex, and unpredictable events, both large and small, on the road. Some end in a crash or worse, but in most cases you can adjust and respond instantly towards a safe outcome.
Unfortunately, algorithms that are not designed to learn from just that kind of incident are not very flexible.
So how do we make sure that self-driving cars are trained on potentially life-saving road-sharing experiences when Mad Max is about to overrun the Fast and the Furious crew? Is not it. Drive Miss Daisy millions of miles around strip malls?
again: ChatGPT has zero intelligence, but it’s a usability revolution, say AI experts
Liu and his team begin collecting real-world data, such as speed and direction, from hundreds of privacy-preserving sensors at smart intersections in Ann Arbor and Detroit that provide a wealth of traffic data, including accidents. Did.
In addition, up to 160 volunteer passenger cars were also equipped according to the study.
One particular two-lane roundabout turned out to be a hot spring for accident money, as roundabouts were largely unfamiliar to American drivers. Liu knows the area well and has repeatedly taken his unlucky son there to acclimatize him for his driving test.
Second, a University of Michigan study removed critical unsafe information from driving data. In other words, it removed all the tedious miles of safe driving between accidents, but left behind what ended up being a fender bender. This data was fed into a neural network used to train self-driving cars.
again: What is deep learning and everything you need to know
Next the team headed to M-City, something of an automotive Truman Show. This is a pioneering fake city environment with traffic lights, pedestrian robots and other vehicles.
“We created a mixed reality test environment,” says Liu. “The AV test vehicle we use is real, but the background vehicle is virtual, so we can train it to create challenging scenarios that rarely occur on the road.”
In this space, test vehicles will encounter more dangerous situations much more often, but it will dramatically reduce the time required to learn how to do the roads, albeit effectively.
again: An AI bot passed the medical school exam, but should it become your doctor?
Liu estimates that training a Waymo-type car could take just a few thousand miles, including various crashes, rather than tens of millions of smooth pavement miles.
Still, experts say computers in AVs lack the quick, intuitive thinking that humans do when faced with complex and unpredictable situations.
I don’t get a better industry barometer for this mindset than Ford and Volkswagen’s decision to exit Argo, a company that hoped to write off billions of dollars and usher in the dream of self-driving cars.
Bad news for the self-driving industry, but good news for humanity trying to stay ahead of the rise of machines.
