The big new idea for making self-driving cars that can go anywhere

The big new idea for making self-driving cars that can go anywhere thumbnail

Four years ago, Alex Kendall sat down in a car on a rural road in the British countryside. He then took his hands off of the wheel. The car was equipped with a few inexpensive cameras and a large neural network. It veered to one side. Kendall grabbed the steering wheel and corrected it when it did. Kendall corrected the problem when the car veered once more. It took less than 20 minutes for the car to learn to stay on the road by itself, he says.

This was the first time reinforcement learning, an AI technique that trains a neural system to perform a task through trial and error, had been used to teach cars to drive on real roads . This was a small step forward in a new direction that startups believe could lead to driverless cars becoming a common reality.

Reinforcement learning has had enormous success producing computer programs that can play video games and Go with superhuman skill; it has even been used to control a nuclear fusion reactor. Driving was considered too difficult. Kendall, founder and CEO at Wayve, a UK-based driverless car company, says that they were laughed at.

Wayve trains its cars in rush hour London. Last year, it showed that it could take a car trained on London streets and have it drive in five different cities–Cambridge (UK), Coventry, Leeds, Liverpool, and Manchester–without additional training. This is something industry leaders like Cruise or Waymo have not been able to do. Wayve, a company that is partnering with Microsoft to train its neural networks on Azure, the tech giant’s cloud-based supercomputer, announced this month.

Investors have sunk more than $100 billion into building cars that can drive by themselves. This is a third of the amount NASA spent to get humans to the moon. Driverless technology is still in the pilot phase, despite over a decade of development and miles of road testing. Kendall says that “we are seeing extraordinary amounts spent to get very limited results.”

That’s why Wayve, as well as other autonomous-vehicle startups such as Ghost and Waabi, both in the US and Autobrains, which is based in Israel are all focusing on AI. They call themselves AV2.0 and believe that cheaper, smarter tech will allow them to overtake the market leaders.

Hype machines

Wayve says it wants to be the first company to deploy driverless cars in 100 cities. Is this just more hype in an industry that has been relying on its own supply for years.

“There is way too much overselling in this field,” says Raquel Urtasun, who led Uber’s self-driving team for four years before leaving to found Waabi in 2021. “There is also a lack recognition of the difficulty of the task in the first instance. But I don’t believe that the mainstream approach to self-driving is going to get us to where we need to be to deploy the technology safely.”

That mainstream approach dates back at least to 2007 and the DARPA Urban Challenge, when six teams of researchers managed to get their robotic vehicles to navigate a small-town mock-up on a disused US Air Force base.

Waymo, Cruise were launched after that success. The robotics approach adopted by the winning teams remained the same. This approach considers perception, decision-making, vehicle control, and vehicle control as separate problems with different modules. Urtasun says that this can make it difficult to build and maintain the system. Errors in one module can cause problems in others. She says, “We need an AI mindset and not a robotics mindset.”

Here’s the new idea. Wayve, Waabi, etc. are creating one large neural network and wiring it together manually. The AI learns to convert input (camera and lidar data about the road ahead), into output (turning or stopping the bike), much like a child learning to ride a bicycle.

Going straight from input to output like this is known as end-to-end learning, and it’s what GPT-3 did for natural-language processing and AlphaZero did for Go and chess. “In the last 10 years it’s caused so many seemingly insolvable problems to get solved,” says Kendall. “End-to end learning has pushed us forward into superhuman capabilities. Driving will be the same .”

Like Wayve Waabi uses end-to-end learnin. However, it isn’t yet using real vehicles. It is currently developing its AI almost entirely inside a highly-realism driving simulator ,, which is controlled by an AI driving instructor. Ghost uses an AI-first approach to build driverless tech that can not only navigate roads but also learns to respond to other drivers.

200,000 small problems

Autobrains is betting on an end-to-end approach too, but does something different with it. Instead of training one large neural network that can handle all the situations a car might encounter it will train many smaller networks, each capable of handling a specific scenario.

” We’re translating the difficult AV problem into hundreds and thousands of smaller AI problems,” said Igal Raichelgauz (CEO). It’s about extracting contextual cues.”

It’s about extracting context cues Autobrains takes sensor data from a vehicle and runs it through an artificial intelligence that matches the scene with one of many possible scenarios: rain or pedestrian crossing, traffic light, bicycle turning left, car behind, etc. By watching a million miles of driving data, Autobrains says, its AI has identified around 200,000 unique scenarios, and the company is training individual neural networks to handle each of them. The firm has been working with car manufacturers to test their technology and just acquired a small fleet. Kendall believes that Autobrains’ approach might be a good fit for advanced driver-assist technology, but he doesn’t see it as an advantage over his own. He says that if they tackled the whole self-driving issue, he would expect them to be just as challenged by the real world.

Cruise control

Either way, should we count on this new wave of firms to chase down the front-runners? Mo ElShenawy (executive vice president of engineering at Cruise) is not convinced. He says, “The current state of the art is not sufficient to take us to the stage that Cruise is at.”

Cruise is one of the leading driverless-car companies in the world. It has been operating a live robotaxi service in San Francisco since November. Although its vehicles are limited in their operation, anyone can hail a car using the Cruise app to have it pull up to the curb without anyone inside. ElShenawy says, “We see a real range of reactions from customers.” It’s exciting .” Cruise built a huge virtual factory to support its software with hundreds of engineers working on various parts of the pipeline. ElShenawy claims that the main modular approach is a benefit because it allows the company to swap in new tech as it happens. He also dismissed the notion that Cruise’s approach wouldn’t work in other cities. He says, “We could have launched somewhere in the suburbs years ago, and that would’ve put us in a corner.” “The reason we chose a complex urban environment such as San Francisco, where we see hundreds of thousand of pedestrians and cyclists–was very deliberate. It forces us to create something that scales .”.

But, before Cruise can drive in a new place, it must first map its streets at centimeter-level detail. These maps are used by most driverless car companies. These maps provide additional information to the vehicle in addition to the raw sensor data. They can include hints such as the location of traffic lights and lane boundaries, or whether curbs are present on a particular street.

These so-called HD maps were created by combining road data from cameras and lidar with satellite imagery. This method has been used to map hundreds of millions of miles in the US, Europe, Asia. Map-making is a never-ending process because road layouts are constantly changing.

Many driverless car companies use HD maps that are created and maintained by specialists firms. Cruise however, creates its own. ElShenawy says, “We can re-create entire cities–all driving conditions, street layouts and everything.”

This gives Cruise an edge over mainstream competitors, but newcomers like Autobrains and Wayve have abandoned HD maps completely. Wayve’s cars are equipped with GPS, but they learn to read the road by using only sensor data. Although it is more difficult, it means that they are not tied down to any particular place.

For Kendall, this is the key to making driverless cars widespread. He says, “We will be slower to get to our first city.” “But once you get to one city, you can scale everywhere

For all the talk, there’s still a lot to do. Cruise’s robotaxis are driving customers around San Francisco with a pay driver, but Wayve–the most advanced and sophisticated of the new crop- has yet to test its cars in a safe environment. Waabi doesn’t even use real cars.

These new AV2.0 companies have a lot of history: end-to–end learning rewrote all the rules for computer vision and natural language processing. Their confidence is not misplaced. Urtasun says, “If everyone goes in the same direction and it is the wrong direction, then we’re not going solve this problem.” “We need a diversity of approaches, because we haven’t seen the solution yet.”

Read More