The controls for fully autonomous trucks must deal with inputs from a number of sensors, analyzing this data and taking actions that enable the vehicle to complete its route safely and efficiently. Design teams are racing to put together the perfect control architecture for their vehicles, while ensuring that these complex real-time systems operate over long vehicle lifetimes.
Thousands of megabytes of data from different combinations of cameras, radars, lidars and ultrasonic sensors flow into controllers that meld this information into a map of the vehicle’s surroundings. Within milliseconds, these controllers must determine adjustments that keep the vehicle on its path and operating safely. The computing challenges are matched by the needs of creating bulletproof systems.
“Sensor fusion is used to create an understandable surrounding, making it possible for the vehicle to conduct the transport mission in a safe way,” said Johan Larsson, director of Autonomous Solutions, Volvo Trucks North America. “Of course, the perception system setup is redundant, with ‘belt and suspenders,’ meaning that you can operate in a fail-safe mode even if one of your system components would fail.”
The controllers that analyze sensor inputs and determine what the vehicle should do need to operate at extremely high speeds, with little margin for delays or failures. Engineers tasked with building control modules also must ensure that components can meet the strict reliability levels of commercial trucks.
“We’re talking about how to make this commercially viable, that is the key to realizing the goal of getting drivers out of the vehicle,” said Chuck Price, TuSimple’s Chief Product Officer. “A lot of companies are building prototypes that aren’t ready for commercial use – things will wear out before they leave the driveway.”
Before autonomy became a focus, many sensors fed data to their own dedicated controllers. That’s now shifted to centralized management, which makes it easier to make decisions based on 360-degree inputs. Combining controls for various subsystems into a single module simplifies the control architecture. It also makes it easier to update software over the long lifetime of a commercial vehicle.
“A key benefit of the new architectures is that we can provide a centralized infrastructure for all electronic control units,” said Martin Schleicher, vice president strategy, Elektrobit. “This centralized function not only increases the vehicle’s safety, but also ensures that the vehicle’s software can be kept up-to-date during its entire lifecycle. The automated vehicle is another device in the IoT.”
Many vendors are planning moves from limited operating domains to more common on-highway driving. More diverse operating realms require more analysis of the surrounding environment and analyzing data from sensors quickly enough to avoid accidents requires hefty amounts of processing power. Development teams are forging partnerships to help them meet these demands while getting to market quickly.
“It is because of this, and the importance of being early in the market, that we have decided to partner up with Nvidia,” Larsson said. “Nvidia has world-class knowledge of artificial intelligence and computing. We will use their hardware, simulation tools and some of their software. Combining this with our world-class knowledge of vehicles and vehicle control is a good base for the development of a high-performing automated driving system.”
Many truck and automotive companies are using Nvidia’s graphical processing units (GPUs), but GPUs are only part of the processing equation. Today’s systems typically include at least one conventional microcontroller that usually handles decision making, among other tasks. And GPU makers are beginning to include microcontrollers, which may reduce the dominance of traditional central processing units (CPUs).
At the same time, field-programmable gate arrays (FPGA)s give design teams the ability to combine conventional CPUs with several parallel processors like those found on GPUs. FPGAs can be adapted to meet different challenges, letting developers adapt configurations to meet specific requirements.
“It’s a challenge to find the right balance between GPU demands and general-purpose computing demands,” Price said. “Today, we’re mixing ruggedized GPU systems along with Intel Xeon processors that run non-GPU algorithms. It’s the role of FPGAs to optimize certain elements of the system.”
These centralized architectures still need to be designed to fit in limited space, dissipate heat and keep wiring harnesses short. At the same time, components must be cost-effective to replace. Modular designs can help engineers meet these requirements.
“Coming from a distributed architecture, we are moving in the direction of a more centralized architecture,” Larsson said. “But we do use a modular approach in the design of the automated-driving system, with clear interfaces to support applications with different requirements.”
Powering all the cameras and chips on a large truck takes a lot of energy. Some processor chips can draw close to 100 W, and most systems have a large number of processors. Cameras also consume lots of power. Price said that electrical power requirements for TuSimple’s autonomous vehicles range from 3-6 kilowatts depending on the size of the truck and version of the system architecture. That puts pressure on the vehicle’s power generators.
“Power is a big deal for us. We’re not compute-constrained, we are power-constrained,” he said. “We consume almost all the power in a Class 8 truck with a big engine and a big alternator. Systems are quite power hungry, there are a lot of cameras.”
While the goal of long-term autonomous programs is to get people out of the vehicle, some human intervention may be necessary when unusual circumstances arise. Truly driverless vehicles will pull over and stop when their controllers can’t navigate through confusing situations. Some suppliers believe that when onboard systems can’t understand tricky situations, help can be provided by remote operators.
Starsky Robotics last year demonstrated its capabilities with a nine-mile drive managed by an operator who sat in a remote data center. Another remote-control proponent, Designated Driver, is demonstrating the technology on buses in Texas. The company thinks that commercial vehicles are an area ripe for expansion.
“Commercial truck and heavy equipment manufacturers are definitely looking at teleoperations as a potential solution for enabling a tighter and younger labor pool, creating work environments that are more comfortable, appealing and safer for a new generation of employees,” said Walter Sullivan, CTO, Designated Driver. “Being able to distribute employee resources across distant geographies for some jobs can be a huge advantage in managing employees, utilization, and service.”
Not everyone agrees with this concept. TuSimple’s Price feels that latency and other issues make it quite difficult for remote operators to make real-time decisions.
“We don’t believe remote driving can be made safe or reliable,” Price said. “We do believe we can remotely alter the plan for the vehicle when it’s in a minimum-risk condition. We can tell it to move forward 20 feet and left 2 feet, and to do that using the vehicle intelligence to determine if there’s anything in its path and make other decisions.”
Borrowing from military, industrial
The sensing and computing requirements of autonomous trucks are high, but volumes are low, forcing design teams to borrow technologies from other fields. Automotive systems are the obvious choice, but as processing requirements rise, military equipment suppliers also are focusing on large commercial vehicles.
Commercial-vehicle developers have long leveraged the technologies and volume pricing of the automotive industry and that model will continue for autonomous technologies. While many automotive-grade boards and sensors are designed into trucks, ruggedization of these components is a necessity because reliability requirements are dramatically different.
“Cars are designed for longevity of 100,000 to 200,000 miles, so that’s what component suppliers design to,” said TuSimple’s Price. “Heavy trucks are specified to a million miles. That changes the nature of how things are built.”
The size of trucks is another major difference from passenger cars. It’s harder to stop large trucks and it’s trickier to maneuver them, especially when traffic gets heavy. Control systems and sensors need to account for these differences, mainly by extending the sensors’ field of view.
“Most of the technologies can be borrowed from passenger cars, but in many cases, there will be specific tailoring to the ‘truck use case.’ One thing that will lead to specific technology development for trucks is the need for a longer range of the perception system,” said Volvo Trucks’ Larsson. “A truck needs to see much longer than a passenger car to be able to secure a safe stopping distance. Another challenge is that some traffic situations, for example, merge scenarios, are more complex with a tractor-trailer than with a passenger car.”
Ruggedization specifications and computing levels for electronics used in commercial vehicles are quite similar to the requirements set for many military components. Connected-vehicle needs are also like those of industrial equipment, which has for years offered connection to the Internet of Things (IoT). Connectivity also is a standard feature on many trucks. Those commonalities are attracting some military system providers.
“Trucking, smart agriculture and mining vehicles have requirements and environments that are similar to defense systems,” said Joe Eicher, director of business development at Kontron. “These applications are also tying into areas like the IoT, where our industrial group plays, so our eyes are squarely on autonomous vehicles in mining, ag and trucks.”Continue reading »