TuSimple Takes a LiDAR-Free Approach to Autonomous Trucks

Ask Xioadi Hou, Co-Founder and CTO of TuSimple, about his startup company's methodology for creating artificial intelligence for autonomous trucks, and one answer quickly jumps out: “We're not using LiDAR,” Hou told Design News. “The biggest difference between us and other companies is we do a lot of camera-based analysis.”

Since its founding in 2015, Hou's company—TuSimple, a Beijing- and San Diego-based startup—has been developing AI for autonomous trucking. This domain of the autonomous vehicle space is populated by only a handful of other companies, such at Otto (now a subsidiary of Uber). The company's name comes from the Chinese character tu, which means “image"—playing up the company's goal to make computer vision and image recognition simplified. But it's also a play on “too simple.” “It's very sarcastic because we're overly optimistic,” Hou joked.

“I wouldn't really consider LiDAR as an evil thing. I'm from the utilitarian point of view, where anything that works should be incorporated into the vehicle,” Hou said. “But, frankly speaking, LiDAR today is not an option for a production level vehicle, which is why we have to work really hard on a camera solution to solve the problem.” And, sarcasm aside, while the autonomous car has yet to be perfected, Hou said that trucking brings its own set of challenges. Speaking with Design News from the 2018 GPU Technology Conference (GTC), Hou and his team readily admit that their shunning of LiDAR tends to raise eyebrows. But they have a good reason behind forgoing the laser sensor technology in favor of camera-based systems.

Hou said TuSimple is open to implementing LiDAR into its solution when the technology is more ready. But he offered that the same issues that make a commercial truck so difficult to drive compared to a car also mean that autonomous systems have to be approached in a different way. “For trucking, you're talking about an 18-wheeler driving 60-65 miles per hour on a highway. At that speed, you just can't stop the vehicle within 100 meters of braking distance,” Hou said. “There are claims that LiDAR can see over 200 meters, but it's not always guaranteed. Sometimes the LiDAR can only see about 100 meters or lower. It's not guaranteed that LiDAR can see everything at 200 meters.”

There are also issues, such as material reflectance, that affect LiDAR's perception quality. Things like a black car or a pedestrian wearing fleece could compromise a LiDAR system. In fact, LiDAR was blamed as the culprit behind a fatal autonomous car accident that happened back in March in which a self-driving car fatally struck a pedestrian on a bike at night.

All Cameras, All the Time

Models of TuSimple's solution involve an array of four to five cameras precisely retrofitted to a Peterbilt truck. “We just can't trust one signal unit, so we have to think about what we need and have a series of different cameras,” Hou said. “We use cameras with different lens sets and orientations and it's sufficient for the entire system to work. We also use radar for backup in severe weather. But relying on radar doesn't support full automation.”

While the cameras used in TuSimple's system are off-the-shelf models with some proprietary modifications, TuSimple is not a hardware company. Rather than build its own truck, the company aims to supply the AI and software for trucking manufacturers to build their own self-driving trucks.

RELATED ARTICLES:

With that, however, comes concerns over meeting manufacturer specifications—particularly around additional power consumption. Hou laughed that TuSimple's engineers are always asking for more power consumption, but understand they have to work within limits. “We currently consume about 2000 watts just for computation. Aside from that, there's very little retrofitting on the truck,” he said. “We don't want to build something in a very hacky way because then it won't be very safe or reliable. What we want is a production level system. In order to do that, we have to buy parts from Tier 1 suppliers because they can guarantee the parts are production-level quality. We don't need to rebuild the power steering system or build our own braking system. All of those components are there. We just need to talk to the right people and collaborate with them in the right way.”

Chuck Price, VP of Product at TuSimple, told Design News that things such as changes to the power system are steps toward working with TuSimple's trucking partner, Peterbilt, to realize a truck design that can come off the assembly line ready for full autonomy. “A large part of our business is providing a reliable autonomous service,” Price said. “We're licensing our technology to OEMs and other Tier 1s that will be building the long-term production platforms that we'll require. It's not our preference to be in the hardware business directly, but to be more in the technology business for that part of the system.”Will the emergence of electric trucks further alter the landscape? Electric trucks manufactured by companies like Tesla seem to be rising right alongside autonomous vehicles. But Hou doesn't believe the two are necessarily related. “People are always talking about electric trucks, but it has little to do with autonomous driving,” he said. “Is it good to have an electric vehicle? Yes, because the response time for an actuation improves. For example, if you want to have more torque on one of the wheels, we can control that more precisely than on a diesel engine. We can also easily draw power from the unit. But those are the only things that I feel will be good to have. Those are perks, but they're not going to change the overall course of autonomous driving. EVs are good to have, but not mandatory for autonomous driving.”

Machine Learning Is 'Deep' Enough

TuSimple's timeline is right on track with analyst predictions that level 4 autonomous vehicles (vehicles capable of full autonomy within a specific driving domain) would be in testing and pilot programs this year. While TuSimple's system equips trucks with level 4 autonomous capability, Hou isn't caught up in the labels and classifications. “Level 4 is what we're currently working on,” he said. “And there's a huge difference between level 3 and level 4, but I do not see a clear gap between 4 and 5. For example, for level 5, you have to remove the steering wheel. But since we don't want to build a vehicle by ourselves, we're fine calling ourselves level 4.”

He added that at the end of the day, TuSimple is about creating functions that serve its clients' business. “Level 4 or 5 is a terminology thing that doesn't really have much impact on us.”

TuSimple's AI is built primarily around machine learning. One would imagine a task as complex as piloting a truck would require significant deep learning capability, but Hou said that machine learning is more than sufficient. “We use a lot of deep learning, but deep is only a fraction of our code base,” he said. “People today are getting high on deep learning, but deep learning is getting shallower and shallower. And machine learning is as deep as before.”

This is also the reason TuSimple has chosen to take on the task of handling the 3D mapping for its autonomous trucks. “The reason we do the mapping by ourselves is, even though our vehicle can run without assistance from the map, we feel we need the reliability that comes with doing precomputations using the map so we can confidently drive on the road without running into situations where the algorithm fails,” Hou said. “An algorithm will always fail. But if we have a precomputed map and know where we are, it will be much easier to correct any failure and save us from any unpredictable circumstances.”

The camera system could generate real-time maps for the trucks and run on uncharted terrain, but Hou said such a system isn't as reliable. “The key to autonomous driving isn't about flashy technology; it's about reliability,” he said. “How can you guarantee your system is running 20 hours a day, six days a week, and you're indestructible? You have to be at that level to make sure this is a viable product.”

A Fully Supported Fleet

To date, TuSimple has done road tests for about 15,000 miles in China and the US combined. The company runs a 50,000-square-foot testing facility in Tuscon, AZ and an equivalent facility in Beijing that is 60,000 square feet. Arizona regulations allow TuSimple to conduct testing as long as human drivers are in place, and the company has access to a 40-mile stretch of highway in China for testing in mixed traffic.

“The benefit of having our own facility is it allows us to test the technology with real loads,” Price said. “It also allows us to test our business model. We're trying to have no negative disruption and very positive disruption in our ability to drive trucks longer; to go slower but get there faster, and these kinds of things.”

Price added that TuSimple plans to offer road assistance and service to clients. “If you get a flat tire when you've got a driver in the truck, you've got a ready-made solution because the driver can fix the tire himself or make a phone call. But when it's an autonomous truck without a driver, what do you do?”

When its systems are deployed in truck fleets, TuSimple will be able to constantly monitor the vehicles and provide some form of rescue service when an issue is detected—whether that's sending out a local repair service or even sending a human driver to the site. Price said this rescue service will be offered via a subscription model that will include remote map and software updates as well.

“The goal for a fleet is to require no additional burden,” he said.

Chris Wiltz is a Senior Editor at Design News covering emerging technologies including AI, VR/AR, and robotics.