Autonomous Vehicle Design Begins To Change Direction

Tools that are commonly used in semiconductor design are starting to be applied at the system level for assisted and autonomous vehicles, setting the stage for more complex simulated scenarios and electronic system design.

Simulation is well understood for designing automotive ICs, but now it also is being used to design vehicle architectures and sensors, as well as for sensor miniaturization and for integration within a vehicle. So instead of just simulating chips, these tools are being used for modeling dynamic behavior and possible interactions of vehicles, which is much faster and more efficient than driving billions of miles to find the corner cases.

“They are modeling things like traffic simulation and managing fleets of vehicles,” said Andrew Macleod, director of automotive marketing at Mentor, a Siemens Business. “And when you’ve got all these different autonomous pods and small batches, can you model the factory? Cloud/data analytics of performance of these vehicles is being fed back into the design and manufacturing.”

Jason Andrews, solutions director, development solutions group at Arm, points to a similar trend. “Autonomous driving relies heavily on the ‘sense, perceive, decide, and actuate’ process. Gathering information from sensors and deciding what action to take is critical. There are multiple approaches to designing systems around these four stages, but there are multiple core options available that play a role in understanding sensor data. The best way to start experimenting with design choices is early software development using simulation models of CPU subsystems and sensor data.”

There are many entrenched processes today for developing automotive systems. Not all of them are suited for the complexities and challenges of autonomous driving, however.

“The key is to consider the V cycle as we know it in the automotive industry as something of the past,” said Matthieu Worm, program lead, autonomous driving and vehicle dynamics at Siemens PLM Software. “We need to move to a vehicle deployment process that takes into account user data that is collected from vehicles on the road, and to be able to update the software that runs in the vehicles that are on the road to improve their performance. If you that that as a given, plus the fact that these cars will have more and more autonomous functions, these are two very strong ingredients to force the industry to rethink the way products are designed. What we consider the main ingredient of the solution is a model-based system engineering approach, where you generate digital twins with a high level of fidelity to assess these software updates and to assess the safety levels of autonomous functions.”

This has broad implications for automotive design, such as the role of sensors in a car and how much intelligence should be put at the edge. “A ‘dumb’ sensor will take a small amount of power and be relatively inexpensive to replace in the case of an accident,” said Rod Watt, director of vehicle architecture and system analysis in Arm’s Embedded and Automotive Line of Business. “However, the lack of pre-processing at the edge will mean more data will need to be transmitted and computed at the center. A more complex sensor, capable of pre-processing, will cost more but ultimately take some of the compute burden from the central processor and reduce the data bandwidth, but valuable data could be lost in the process. In all scenarios, safety, security, and getting the correct amount of processing in the right place in the chain is key to an optimal solution.”

This can vary greatly by region, too. While Europe and North American designers favor more localized processing of data, China is pushing for more centralized processing over high-speed communications rather than building those capabilities into the vehicle. That alters the basic design of the electronics, but it doesn’t alter the complexity or the amount of data that needs to be processed, which is enormous.

“Humans are not able to read data in more than five columns,” said Burkhard Huhnke, vice president of automotive strategy at Synopsys. “You need a computer to do that. It takes a system to observe and monitor all the functions of an autonomous car, and you still need some intelligence built into the car.”

Huhnke noted that one of the most effective ways to achieve this is with pre-developed IP and a standardized infrastructure. “You need to begin standardizing the infrastructure. Autonomous cars will require 150,000 to 200,000 lines of code. Now one can afford to do that anymore. Even at large automotive groups they are beginning to standardize on open infrastructures to provide over-the-air updates.”

The rising value of simulation
While standardization helps with real-world testing, assessing the nominal performance of individual sensors and combinations of sensors under all possible real-world scenarios is impossible. “As a result, we’ll need new technologies and approaches to assess performance, and simulation is a very important one,” said Kurt Shuler, vice president of marketing at Arteris IP. “This is a new area, and the industry is just starting to understand the magnitude of the issues. But the new ISO PAS 21448 SOTIF spec should be helpful in providing a common framework and terminology to start this important work.”

Additionally, because the key sensor functionality gets more and more realized in silicon systems, sensors must be simulated to ensure that the specification is met before they get implemented in silicon, reminded Robert Schweiger, director of automotive solutions at Cadence. “A perfect example is LiDAR, which is currently transformed from a macro-mechanical scanning LiDAR system to a highly-integrated solid-state LiDAR sensor in one example. A solid-state LiDAR sensor blends multiple chips from different silicon process technologies mounted on varying substrates and housed in a variety of package types. [Commercial tools] integrate IC design and package analysis to enable concurrent design across chip, package, and board. Traditional chip design tools, however, do not accommodate the effect of packaging and require manual corrections when validating design performance.”

Simulation for these kinds of multi-technology systems is required to ensure that the laser beam is correctly and precisely controlled by the MEMS mirror, he stressed. “Here, the silicon photonics detector, realized as an array of single-photon avalanche diodes (SPADs) or silicon photon multipliers (SPMs), needs to be carefully designed to ensure the highest sensitivity. That determines the range of the LiDAR system as one of the key performance indicators of such a sensor. An integrated electronic/photonic/MEMS system simulation is key to ensure the overall system performance.”

Interestingly, formal verification is playing an increasingly important role in automotive system design as sensors become a main safety component of autonomous vehicles.

“High-level scenario modeling using ESL or portable stimulus is valuable, but it is unlikely that these techniques will exercise deep corner cases within the sensor processing logic,” said Vladislav Palfy, director application engineering at OneSpin Solutions. “Formal verification is ideal for hitting corner-case conditions and detecting any associated bugs. Multi-sensor systems are inherently complex, with many possible combinations of data, making it impossible for simulation to iterate through all the combinations. Only the exhaustive power of formal verification can do the job completely and efficiently. Finally, safety verification is required where a random error in the field could cause an incorrect result of consequence, such as a proximity sensor with a false reading leading to a dangerous and unnecessary evasive maneuver. Again, formal verification should be the tool of choice.”

For automotive OEMs, this is a time of huge change with the move to autonomous vehicles.

More collaboration required
Siemens’ Worm observed there is currently a lot of tension across the supply chain, especially between the OEMS and the silicon vendors. He believes that what is currently preventing big progress is the discussion on IP and how to share models between silicon suppliers, Tier 1s, and OEMs without giving away the IP. “We need to have an opportunity to share the virtual representations of sensors, the electronics but also the sensors themselves could be mounted on virtual vehicles to do a full vehicle assessment.”

Currently, Worm noted, the OEMs reverse engineer sensors or equipment they get from suppliers to do a virtual assessment. “That’s very time consuming and far from an ideal representation of the actual hardware.”

This really speaks to the recurring theme of collaboration in the supply chain, according to Macleod. The silicon vendors are looking, in many cases, to extend their technological advances into automotive, so there are a lot of new entrants as well as the established players in the supply chain trying to work out such things as how to get ISO certification, how to make sure these vehicles and components are safe, and at the sensor level, how to make sure the concept of security is understood.

The semiconductor houses also are looking at how to help solve some of these high-value problems, such as how to facilitate innovation, and a lot of that involves adding software layers onto the silicon.

“This whole concept is helping to pave the way for new automotive business models that carmakers are not going to be selling physical cars exclusively,” Macleod said. “They’ll be selling trips and all kinds of other things. The semi houses are trying to help the other tiers within the concept of collaboration. Carmakers understand very well, especially with autonomous, that liability is going to be stopping with them, so this innovation and getting to market fast is one aspect. But how do they make sure these things are safe? There must be visibility throughout the whole supply chain. If they are designing bits of IP themselves, passing RTL onto the silicon suppliers, how do they make sure it is to spec and delivered on time. There’s a whole bunch of new challenges that fall under the collaborative supply chain model that we didn’t see 10 years ago.”