TuSimple Uses AI to Train Self-Driving Semis

Using AI to Make Fully Autonomous Trucks

For all the hoopla surrounding self-driving passenger cars, there’s actually a more pertinent problem that autonomous vehicle technologists are working to solve: bringing self-driving technology to long-haul trucks.

Trucking isn’t cheap. It costs an average of $1.69 per mile to operate a long-haul truck, and about 40 percent of that represents the driver’s salary. Meanwhile, truck drivers are hard to find, with one estimate suggesting that the U.S. freight industry currently faces a shortage of 100,000 drivers—and that could triple by 2026 as e-commerce continues to push the demand for freight higher.

Trucking is also dangerous. Large trucks were involved in more than 4,700 fatalities in 2017, an increase of nine percent over 2016, despite an overall decline in passenger vehicle deaths. As a result, “truck driver” is now classified as the most dangerous job in America.

One solution to these challenges is clear, if complicated: have some trucks drive themselves, leveraging artificial intelligence (AI) to fill the existing driver shortage and increase safety levels. But building self-driving trucks is no trivial task.

Making Trucks Fully Autonomous

TuSimple is at the forefront of this effort. The company, based in both San Diego and China, already uses self-driving trucks to make commercial deliveries, primarily in Arizona.

Founded in 2015 by President & CTO Xiaodi Hou, TuSimple now employs about 400 people between its two headquarters and makes three to five delivery trips per day, each one about 100 miles in length. These trips are largely used to generate testing data for TuSimple’s deep learning algorithms, which are constantly working to master the rules of the road, but they’re also making money.

“There are many companies that have vehicles on the road,” says Hou of TuSimple’s competitors, “but not many who are actually hauling real commercial cargo.”

At present, TuSimple’s trucks are running at a “level 4” autonomous vehicle classification. The trucks fully drive themselves, but a human driver is always in the cab, ready to intervene if things go awry. Hou wants to remove the human fail-safe by the end of 2020, and he says the company is on track to make this happen.

“There are many companies that have vehicles on the road, but not many who are actually hauling real commercial cargo.”

Xiaodi Hou
President & CTO
TuSimple

“There are many companies that have vehicles on the road, but not many who are actually hauling real commercial cargo.”

Xiaodi Hou
President & CTO
TuSimple

AI in the Driver’s Seat

Getting a tractor trailer to drive itself for hundreds of miles is no easy feat, and the two biggest hurdles should be clear to anyone who’s ever shared a freeway lane with a big rig. First, trucks are very long. While a self-driving Prius only has to navigate 15 feet of length, a full tractor trailer is about 70 feet long, making every calculation about accelerating, stopping, and merging more complex. Perhaps an even bigger challenge, says Hou, is the width of the truck. A passenger vehicle may not even occupy half the width of an Interstate freeway lane (in the U.S., 12 feet wide by law), while a tractor trailer typically has only 21 inches of room on either side of the freeway lane—and much less on a city street.

The calculations surrounding these vehicles are incredibly intense, and they have to be done in motion and on the fly, with human lives at stake if something goes wrong. The hard work is done courtesy of a computer—Hou calls it a “supercomputer”—installed directly on the truck. Each TuSimple server is loaded with up to 100 different AI modules tasked with doing everything from distinguishing the types of cars on the road to determining the speed of other objects around the truck (the toughest algorithmic task of the bunch). Built primarily using the Apache MXNet deep learning framework on Amazon Web Services (AWS), these modules suck in a steady stream of data from the system that uploads up to a terabyte of data from cameras, LiDAR, and radar equipment mounted on the vehicle, that then use that data to build a live 3D model of the road that is constantly updated as the truck continues down the road.

“The system also has to work in all conditions, not just the best conditions,” Hou says. “Say you’re driving at dawn or dusk. How do you deal with that? What about rainy situations?”

The interaction among all the sensors, the algorithms that analyze their data, and those that actually control the truck is extremely complicated, and it’s here that advanced AI comes into play most clearly. Managing all of these data points and using probabilistic analysis to make the best possible decision several times each second is one of the most complex computing tasks around—and a machine learning environment is the only ecosystem that can handle it. And, according to Hou, the technology is extremely accurate: At 65 mph with a loaded trailer, the control algorithm can keep the truck at the center of a lane with 5cm accuracy.

Ready to Ride Another Day

After a successful delivery, TuSimple’s trucks return to their home base, and the work of analyzing all the data collected on that run begins. The results are updated to the algorithmic modules on every truck’s onboard server. That may sound simple enough, but every time an algorithm is updated, it has to be run through thousands of safety tests and simulations to ensure it will behave as expected on the road.

Hou says that running these tests in-house would take weeks each time new code is rolled out, but thanks to the massive computing power of AWS GPU-powered compute in Amazon EC2 P3 instances, the tests can be completed in a matter of hours, minimizing the turnaround time before a truck can get back on the road. TuSimple is also currently evaluating AWS Snowball Edge, a new edge computing device that would streamline the migration of all of this data from the trucks to AWS cloud services and back.

Looking ahead, Hou says the day when most trucks are operated by computers is not something to be scared of: It’s inevitable.

“A few hundred years ago we were using cows to help us plow the land,” says Hou. “We’re still looking for ways for human labor to be done by machine. This is really a natural evolution of human civilization. I’m very proud I’m on the forefront of it.”

Zocdoc uses AI to build patient confidence

Learn more »

Celgene uses ML to transform drug discovery

Learn more »

GE Healthcare drives better healthcare outcomes with ML

Learn more »

ML helps Duolingo personalize language learning

Learn more »