Making Money with Self-Driving Trucks? It’s TuSimple

You know those annoying drivers who merge onto the freeway without so much as glance over their shoulders? For you, a driver already on the freeway, it can be infuriating to encounter a vehicle that has forced its way into your lane. You may have to brake, swerve, speed up… you might panic. But watch a long-haul trucker in a similar situation and you’ll see the grace with which the situation can be managed. The driver will smoothly blend into the lane to his or her left —long before it’s necessary—then blend back in when the time is right. There is no braking, no honking, no sweat. The annoying driver may never have noticed the big rig as it rolls by.

That’s how the pros do it. Taking lessons are self-driving vehicle companies, like TuSimple, whose vehicles are the trucks that will soon to be driverless. But as we are about to see, programming for the annoying-driver-merging problem is anything but too simple.

It Only Looks Simple

There’s a lot for an autonomous vehicle to consider, according to TuSimple's founder and CTO, Xiaodi Hou, who was at NVIDIA’s annual geekfest, the GTC conference, once again this year, this time expounding on “AI – the Good, the Bad and the Ugly.”

• You have to recognize the moving objects as vehicles.
• The vehicle on the right in the merge lane—will its path and speed make it intersect my path?
• Should you slow down and let it go in front of you?
• If so, will you fit after it, or is there another car after it (Car B)?
• But braking is bad for gas mileage….
• Is it better to drift over one lane to the left?
• Is there space between vehicles in the lane to the left (Car C and D)?
• If not now, can I speed up or slow down into the next slot (after car D)?

That’s just a start. There are more possibilities, more potential moving parts and more outcomes. The decision tree branches quickly and often. For an AI system, something that a professional driver does with ease turns out to be a remarkably complex maneuver.

What Is TuSimple?

It could take a Ph.D. from CalTech to figure out just that one merge scenario. Luckily, Xiaodi Hou has one. He earned it in the field of computer vision and machine learning.

We hear TuSimple uses “all the sensors” on its trucks, like the Peterbilt 579 on display. They have to. “LiDAR doesn’t work in the rain.” Trucks have it hard—forced to fit in lanes only inches wider than they are. Plus, they are “230 percent the length of a Camry.” It’s simply much harder for them to change lanes. There’s no place for the truck to go.

Who knew a Camry would terrify a truck driver?

Why Trucks?

Hou gets this a lot. “That's where the money is,” he answers quickly. As much as we all think self-driving car features like automatic braking are necessities and question why they haven't already been mandated, the passenger car industry is faced with serious economic considerations. It costs a lot of money to completely outfit a car with the type of computing and sensors that are needed for Level 4 autonomy. A full rack of sensors (radar, cameras, LiDAR) on the roof and supercomputer in the trunk will easily add tens of thousands of dollars to the cost of a car. Automakers know the typical car buyer will not stand for it.

Trucks are a different story. Their AI systems can be more simply architected because there are fewer scenarios and fewer threats for them to sort out. The expressway, where a long-haul truck logs most of its miles, is a relatively controlled environment. There's nobody zoned out on a cell phone crossing the highway, no kids chasing ball or dogs chasing kids, no traffic lights, no bicycles. Also, the high cost of trucking (the new Peterbilt 579 EPIQ semi truck that TuSimple parked in the exhibition hall costs about $180,000 new). The high cost of a truck makes the added cost of autonomous vehicle “accessories” easier to swallow.

Then there's the operational costs of trucking. There’s lots of money to be saved with autonomous driving.

“Fuel is 20 percent of the costs of a trucking operation,” we are told. Those big trucks have engines the size of cars and get between 5 and 6 miles per gallon. TuSimple hopes trucking fleets will consider that AV software can be programmed to save fuel. That’s second in priority to safety, of course, adds the company. TuSimple makes choices to change lanes rather than brake because braking wastes energy, said Hou. “We can save 20-30 percent in fuel costs with our vehicles.”

In addition to money being saved, there’s money to be made. Unlike self-driving cars, self-driving trucks can carry loads for money. Ride-sharing services ferrying passengers, such as in conjunction with Lyft in Las Vegas, are not allowed to charge for the ride. Trucks can—and do charge for freight, offsetting some of the costs of testing.

Trucking costs are $1.8 per mile. We can recoup as much as $1.6 per mile from carrying freight, explained Hou, reducing operational costs of the beta to 20 cents per mile.

With the safety records of self-driving car companies, can they be trusted? By law, AVs have to report incidents, such as the number of times an operator had to override the self-driving system. While that was meant to be a sign of AV’s requiring intervention because they would be unsafe, Hou claims that AV companies interpret the results in their favor. He is lobbying for danger classifications that are more meaningful that the industry would be accountable for:

• Critical are incidents that result in an accident.
• Traffic violations are costly, too. You can get a ticket for failure to yield or going too slowly, or stopping on highways.
• Most benign are interventions that cause vehicles to slow down, go around vehicles, or stop on the roadside.

“A superior pilot uses his superior judgment to avoid situations which require the use of his superior skill.”

Hou quotes Apollo astronaut Frank Borman (who later became CEO of Eastern Airlines) to explain TuSimple’s forte: using AI to predict how vehicles will act. For example, the merging vehicles will probably have inattentive drivers—so better just move over one lane in anticipation.

The superior pilot in this case was he AI on a collection of computers behind the curtain, stuffed with NVIDIA GPUs, where the sleeper section would normally be, that was shown to us by Chuck Price, chief product officer at TuSimple, who invited us into the cab for an explanation. The AI is able to handle more moving objects and make decisions faster than human pilots. The trick is to be able to recognize far away objects from optical images up to 1,000 meters away, said Price. LiDAR is only usable up to 200 meters, so machine vision is critical to trucks, added Hou.

All sensor-generated data (like video and radar) is mapped onto an object for which the behavior can be predicted. For example, consider what may be an AV’s most formidable problem: a point cloud maps to a small child. Now, what? The kid could dart any which way, a maximum number of degrees of freedom, a nightmare scenario from any aspect. Another set of points is labeled “person with a phone.” If they are approaching a curb, it is likely they will not turn their head before stepping off and detect an oncoming vehicle.

A truck on a highway is only a subset of all the traffic situations possible, but still provides plenty to keep the CPUs and GPUs overheated. Another truck identified and moving on a perpendicular path may likely stay on that road, passing overhead on an overpass: it presents no threat. But that car that is fast running out of a merge lane may soon be under your wheels.

Truckload of Data

TuSimple strives to reduce “false positives,” the system’s identification of threats that really are not. It just makes the big rigs brake and swerve, wasting fuel.

Tu Simple’s vehicles generate truckloads of data. Every ride is closely monitored, and the data produced is carefully studied, says Hou, all in the hopes of making for a smooth (read fuel saving) and safer trip.

A loaded tractor trailer, with its mass, needs to sense farther ahead than lighter vehicles. Image sensors can see far. “Our vision can see galaxies,” said Hou. But we only need to see and react to objects up to a thousand meters away (about two-thirds of a mile).

The Uber Effect

We were introduced to TuSimple at last year’s GTC, not long after an Uber AV had mowed down Elaine Herzberg, a homeless woman who was walking her bike across a road at night near Tempe, Ariz. The AV system had failed to recognize Herzberg and her bike. The Uber vehicle, a Volvo SUV with a sensor rack designed by Uber, had its system tuned for a smoother ride that would decrease its sensitivity to threats and reduce braking or evasive maneuver, reported ArsTechnica,. But even if the system had detected Herzberg, the vehicle was not equipped to apply the brakes. For that, it relied on its operator. Uber had taken to hiring part-time operators for its AVs and stopped having their own “safety officers” accompanying them. Too bad that on that dark night, the operator was watching the “The Voice” on her cell phone, according to local authorities.

It was a tragedy in which many actors played a part. It cast a pall over GTC18 and sent shock waves into an AV industry that had been high on being the next big thing, with hype, startups and investments. That one traffic fatality caused by an AV (vehicles overall caused 40,000 fatalities in the U.S. in 2018) put the brakes on Uber’s self-driving fleet in Arizona. The year before, the governor had welcomed Uber with open arms.

How did TuSimple react to the Uber fatality?

“We don’t allow our operators to have their mobile devices in the cab,” said TuSimple’s Price. “We have two people in the cab. One is in the operator, in the driver side, hands near the wheel. The other, beside him, is monitoring the truck’s performance—as well as the operator’s.

TuSimple operates its self-driving trucks in Arizona. The company seems to know that if a self-driving car can cause public fear and a government backlash, it would be much worse with a self-driving truck. An 80,000 lb vehicle traveling at highway speeds, seen in the rear view mirror with no driver can be disconcerting enough. But an accident? A death? That cannot happen.

It’s not fair to grill TuSimple, to hold it accountable for the sins of Uber, to demand that there never be an accident -- ever. They are dealing with massive, fast-moving vehicles with enormous destructive potential. If the AV industry gets it right, the number of accidents and fatalities will plunge. The insurance industry is already worried that business will dry up as being in a vehicle becomes less risky—and there is no driver to insure. With the total life-saving potential of AV technology, is it fair to burden the AV industry with the expectation of a perfect record, a 100 percent no-kill demand, to insist not a single life can be lost -- regardless of saving tens of thousands of lives?

Price, in his measured and thoughtful responses, was reassuring that safety is, and always has been, very much a concern—the bedrock on which his company rests and builds on, the prerequisite before its AI and technology.