Su Jing (苏箐), Director of Huawei Autonomous Driving Solutions (ADS), gave an extended media interview on April 16 as the company unveiled its smart driving platform run on the Chinese OEM BAIC Group’s new model Arcfox Alpha-S. Su talked candidly about Huawei’s production plan, its take on the industry’s technical debates, and the company’s vision in the autonomous driving space. Below is my summary:

Production

Timeline

• Mass production with BAIC Group at the end of 2021; available for immediate purchase this November or December in Beijing, Shanghai, Guangzhou, and Shenzhen, then expanding to six additional cities every quarter
• Additional models with BAIC Group and other Chinese OEMs (Chang’an, GAC Group) coming out in 2022 – “You’ll see many models coming on to the market next year”
• Huawei is partnering with many OEMs around the world, but the Chinese products come out first because the Chinese OEMs move faster
• Arcfox has been in development for three years; future generations will take two or less
• Software is updated every two to three months; sensors are updated every 18 months

User Experience

• Huawei strives to provide an L4 experience, but legally it’s still L2 (the driver being legally responsible)
• Huawei’s platform offers three modes:
  • NCA (pre-loaded with high-res maps)
  • ICA (what Teslas are like now)
  • ICA+ (self-taught learning – the more you drive in the same area, the better the performance will be)
    • Technically this mode is never going to be able to take the passenger from Point A to Point B, but if they commute between the same few points every day, ICA+ could get a high resolution view of these places and take them from Point A to Point B

Testing

• Testing is done in Beijing (only outside of the fifth-ring road, due to legal requirements), Shanghai, Guangzhou, and Shenzhen, plus major highways around the country
• Testing will be expanded to second-tier cities in the second half of 2021

Edge Cases

• Huawei’s system is more conservative in extreme weather and at night, but these edge cases don’t pose a serious challenge

Business

Huawei’s Philosophy towards Autonomous Driving

• Traditional car companies want to put a new component into the car each time they want to introduce a new feature. But Huawei sees the car as another extension of the computer, rather than the other way around. “Once we build a giant computer, we can put a car in it… That’s what Tesla taught us.”
• Huawei is not interested in manufacturing themselves because the market is bigger if they provide autonomous solutions to OEMs

Huawei ADS’s Organization

• 2,000 people on the ADS team, including 1,200 in perception (divided into vision, lidar, and no further), 200 - 300 in prediction, and 200 - 300 in planning and control
• Core R&D team all in China
• Huawei spent 1 billion USD on the ADS team this year and will probably spend 30% more every year

Huawei’s Business Model

• Huawei will offer both a one-time purchase solution and a subscription model. Subscription fees will be split with the OEMs
• Huawei ADS aims to be profitable in ten years

Differentiation

• Huawei doesn’t look for differentiation in its collaboration with different OEMs – “What kind of differentiation do smartphones have now? If you’re building a product with tens of billions in R&D cost, you shouldn’t be thinking about differentiation in that way.”

Working with OEMs

• Putting a tower on top of a car would make the perception problem a lot easier to solve, but the OEMs taught him that it’s important to make a car look like a car is important
• Many OEMs are still of the mindset that they are putting a computer in a car (rather than building a giant computer).
• Among the Europe auto companies, he’s personally most optimistic about Volkswagen because Volkswagen started exploring autonomous driving early on

Robotaxis

• All robotaxi companies are doomed to fail because they’d need to solve every edge case
• Robotaxis should be the natural outcome of several million autonomous vehicles running for many years, not a business goal to be pursued. “The robotaxi market will be ours eventually, but not now.”
• Robotaxis make even less sense in China because the ride-hailing experience in China is much superior than that in the U.S. and much more affordable, so robotaxis won’t fundamentally change the user experience for people in China

Regulation

• Regulation is not the bottleneck right now; the technology is
• Even though Huawei can offer an L4 experience now, he would not want the driver to leave his seat. Unless the MPI (miles per intervention) goes up by orders of magnitude, there is no truly autonomous driving
• Chinese regulators are very supportive of the AV industry; we shouldn’t blame regulations for everything

Technology

Sensor Set-up

• Arcfox has:
  • three lidar sensors (that will last ten years)
  • four long-focus, wide-angle, stereo cameras, including a set of fish-eye cameras
• The innovation is in the stereo cameras: They are much more stable and generalizable than existing solutions and can see far beyond 20-30m, which is the industry standard today

NCA

• NCA is further broken down into two parts
  • Roadcode HD: offline, high-res maps collected by a professional maps team
    • Shortcoming: new infrastructure gets built in China every day, so the maps go out of date very quickly
  • Roadcode RT: self-taught learning that trains on past data
  • DDI: learns from the driver’s behavior, similar to Tesla’s shadow mode

Vision Only vs. Multi-Sensor

• Data needs to be complex and of high enough dimension to be useful. Tesla’s data is probably no longer improving performance because it’s from vision alone and low-dimensional

Early vs. Late Fusion

• Huawei firmly believes in early fusion and gave up on late fusion two years ago. “Companies that talk about late fusion have their scripts written up by the marketing team, not the R&D team. If you want to improve perception, you do early fusion rather than banking on redundancy. Redundancy is a waste of sensor data.”
• In the demo, Arcfox had trouble seeing a delivery driver cutting in from the back on his electric scooter. This blindspot of lidar sensors is complemented by vision

Data as a Bottleneck

• Data is not the bottleneck now. Algorithms are.
  • Huawei has access to Tier 1 data from OEMs thanks to their partnerships
  • Millimeter wave data is messy, but Huawei is using neural nets to process them

Perception, Prediction, and Planning and Control

• Everyone in the industry knows that perception is difficult, but in terms of technical development, prediction as well as planning and control lag far behind perception. “What a lot of people don’t realize… is that planning and control has a greater impact on MPI than perception.”

MPI as a Metric

• MPI is not a useful metric because it’s highly dependent on the statistical method used, and the time and place the data is collected. “If you deploy several hundred AVs in California at the right time, you can cherry-pick your data to achieve a high MPI.”
• Only MPIs calculated using all historical data is useful. But that metric is a closely held secret at all AV companies. “Huawei can achieve 1,000 MPI in Shanghai [if we cherry-pick our data], but in reality we are not there yet. I don’t think anyone in the industry can achieve 1,000 MPI [with cumulative historical data], not even Waymo.”