Getting Driverless Trucks onto Roadways

Autonomous developers at TuSimple address many technical issues, but they also must consider regulations and operating modes.

TuSimple’s fully autonomous trucks employ more than one type of artificial intelligence to improve decision-making. (TuSimple)

The developers who are creating autonomous systems have plenty of complex technical questions that must be analyzed and solved, but their challenges don’t stop there. Design teams also have to plan for the use cases of driverless vehicles, even going so far as to plan for what might happen if the vehicle has to sacrifice itself to avoid a serious accident.

TuSimple tests its hub-to-hub xe "SAE"SAE L4 self-driving trucks in Arizona, this particular run occurring in heavy rain with zero disengagements. (TuSimple)

The challenges associated with the sea change wrought by autonomous trucking has opened the door for startups like TuSimple, a Chinese startup that’s partially funded by Nvidia. The company is developing digital control systems, focusing on both the prototypes needed to prove the safety of driverless trucks as well as the many issues associated with getting them into day-to-day operations.

“Until you have hardware that’s compliant with safety standards and common automotive standards for survival in harsh environments, plus supplies for 10 years, you’re not really ready,” Chuck Price, Chief Product Officer at TuSimple, told Truck & Off Highway Engineering. “When you’re getting suppliers to agree, it’s one thing to supply prototype volumes, it’s another to get to production. That’s a real challenge.”

Pathway to production

Price and company, which has a U.S. headquarters in San Diego, are also looking at the bigger picture of getting production vehicles on the highways. Some companies have proposed using autonomous controls only on highways, guiding driverless trucks into transfer hubs where drivers would take over for more intricate aspects of delivery. Price contends that this approach has several issues that make it impractical.

Chuck Price (TuSimple)

“We’ve got pretty definitive views how autonomous trucks will be integrated into the transportation environment,” he explained. “Using transportation hubs means you’re only using autonomy on highways, moving into a manual mode for off-highway deliveries. We feel building transfer hubs brings major requirements for zoning and buying land. You’re talking about hundreds of thousands of trucks going in and out.”

TuSimple therefore feels that using autonomous technologies through the whole loading to unloading process is a more preferable approach. Driving autonomously on both highways and city streets, going from depot to depot using automated systems, is the most efficient way to move freight. TuSimple is addressing that by developing sensors and controls that are now being tested on roadways. Real-world tests are important, but virtual testing is also a necessity.

“Simulation is an integral part of the development program,” Price said. “It allows us to do experiments very quickly in a very accurate way. It also lets us do things that are difficult or dangerous to do on highways.”

A new camera system from TuSimple, featuring an advanced automotive CMOS image sensor from xe "Sony Semiconductor"Sony Semiconductor, enhances night vision capabilities to increase self-driving truck utilization from 50% to over 80%. (TuSimple)

These simulations and validation runs will increasingly utilize artificial intelligence. AI helps vehicle controls identify objects and take action in response to their location. Though AI is generally accepted as one of the technologies needed to teach autonomous systems how to respond to the complex nuances of navigating along roadways, critics raise issues.

Detractors note that it’s difficult to verify the accuracy of software responses that change depending on circumstances. TuSimple is addressing that by using more than one type of AI.

“There are different forms of artificial intelligence,” Price said. “Some people are trapped into just thinking about neural networks. We use convolutional neural networks for perception only, using it on sensor feeds from cameras, radar and lidar. When you get into prediction and planning, statistical AI is better than neural networks. It’s more readily characterized.”

That said, anything handled by this automated software will be checked by humans before it’s implemented into the production system. Risk analysis and mitigation will be a major element in all autonomous programs.

“We understand that errors come from AI, it’s not perfect,” Price said. “Companies need to develop mitigation techniques that help assure that the AI is evoking positive results within the bounds of what’s safe.”

Evolving regulatory landscape

Regardless of whether software is written by humans or AI technology, programs will have to meet a range of different regulatory requirements. Those rules also address two different aspects of driverless vehicles, proof of concept and full volume production.

During the current phase of proving the safety and integrity of autonomous technologies, shadow drivers must be present. Some regulators are now considering a change that would allow remote control of vehicles, with humans monitoring them from afar. Eventually, humans could be removed from this loop. However, regulatory changes don’t come swiftly.

“Today’s regulations require a human in the vehicle, but regulations are being rewritten to let systems substitute for the human,” Price said. “To get through that process is the first step.”

Eliminating drivers is also beneficial if a disastrous situation arises. Eventually, an autonomous truck will have to make decisions, such as whether to hit a pedestrian or crashing into a brick wall.

“Unfortunately, it’s possible that a situation may pop up that could be life threatening,” Price said. “If the vehicle is driverless, one option would be to sacrifice the vehicle rather than harm someone or cause greater damage. If there’s a driver in the vehicle, you don’t have that option. We like the idea of taking the driver completely out of the vehicle for safety and efficiency reasons.”

When prototypes evolve into production vehicles, the regulatory environment will also have to advance. While many existing rules of the road will remain in place, a number of new rules will have to emerge.

Regulations will have to address many facets of driving, including breakdowns and accidents. For example, first responders need to know that it’s safe to step in front of an inactive autonomous vehicle that could begin moving.

“There are a lot of practical issues to work through,” Price said. “There must be ways to indicate that the vehicle is in its autonomous stage, especially for first responders who need to know how to approach the vehicle. They need to know whether or not it’s autonomously operating, and how to turn it off.”