Recently I participated in AV19 Silicon Valley, the Autonomous Vehicles conference held annually on the west coast. As the moderator of the closing panel on Industry Standards and Governmental Regulations, I had a chance to pose questions on the future of safety to Walter Stockwell, Director of Safety Policy at Starsky Robotics and Bert Kaufman, Head of Corporate and Regulatory Affairs at Zoox. And while the 3 days of Workshops, Presentations and Panel Discussions clearly proved that self-driving vehicles will save countless lives, the question of “overall safety” kept popping up in my mind. More on that point later, here are a few things that were presented that may be of interest.
The conference got off to a bang as ADAS, self-driving and autonomous vehicle systems will require “massive amounts of data”. Dr. Florian Baumann, CTO for Unstructured Data Automotive & AI at Dell put the data requirements into perspective with some excellent slides. PSA’s Traffic Jam Chaffeur, a Level 3 system needs about 5-10 PB of storage capacity per car. For those who do not know what a PB (petabyte) is, it is 1 MILLION GB. In other words, that is “745 million floppy disks or 1.5 million CD-ROM discs”. What does it cost to store a PB of data? According to this article from 2 years ago, pricing had recently dropped to $10,000/month. I am sure it is cheaper now, but that is a LOT of cost.
The amount of data that is generated and needed by self-driving vehicles is mind numbing. Matthew Park, Head of Business at Scale.ai commented “You need massive amounts of data for self-driving, and no one is talking about it” Google required a billion images to do house number detection accurately.” For those who are not familiar, Google automated the process of recognizing house numbers with AI. The process is limited to 5 digits, and while it took a billion images to develop the software, the initial training took 200 thousand images and 6 days to train the software. While the results were impressive, they would be considered unacceptable for automotive safety; 95% capability and 98% accuracy. Estimates to accomplish auto-safety level software training exceed 10 billion images. And this does not consider new case scenarios like the recent advent of end-mile scooters or special occasions like Halloween (can you imagine trying to train a car to recognize some costumes as “humans”?)
THE BUSINESS OF AUTONOMOUS
There are three major things Americans borrow money for; a house, their education and a new car. Many believe the current trajectory of new car prices is unsustainable (average annual increase is on the order of 3%) and we have not yet fully implemented Level 3 Autonomous Systems (the new Audi A8 was the first vehicle with Level 3 capability but it is not yet on sale in North America). With an automobile being such a large outlay of capital, the advent of ride sharing services seems logical in retrospect. But sharing per Uber, Lyft and others is based on sharing personal vehicles with others. Mark Crawford, Chief Engineer for Autonomous Driving Systems at Great Wall Motor Co., Ltd., presented a MAVS concept (Modular Autonomous Vehicle System) which featured a standardized platform and interchangeable pods/modules that are specifically designed for varied needs. An example presented was a pod that transported people by day being swapped for a pod for cargo delivery at night.
Dr Karl-Thomas Neumann, in charge of Mobility at Evelozcity commented they “will never sell a vehicle”. Their vehicles will be leased and rented and always owned and operated by Evelozcity. This obviously will provide opportunities for round-the-clock use. Ganesh Iyerm, CIO of NIO talked about the full infrastructure of their brand where the vehicle is a part instead of the main component. In China they have battery swapping stations, a NIO app and even a NIO house.
PREDICTING SAFETY OF SELF DRIVING SYSTEMS
The need for self driving is obvious, as Mark Rosekind, Chief Safety & Information Officer at Zoox drilled home. Rosekind took the conference through a substantial number of statistics along with details of significant accidents. Yet consumer support for safety is unclear – how many people would take a 4-star flight yet buying a 4-star crashworthy vehicle is commonplace. More than 34 thousand deaths a year could be eliminated if self-driving vehicles either conquer the marketplace or become government mandated because the 94% that are caused by human choice and distraction could be eliminated.
Perhaps the best presentation I have personally seen over all my self driving conferences was given by Amitai Bin-Nun, Vice President for Autonomous Vehicles and Mobility Innovation at the SAFE Autonomous Vehicle Initiative. Bin-Nun’s presentation centered on how to predict the safety of a particular self-driving system in advance of waiting for years of on-vehicle use data. The material started with the question of ‘how do we know which humans are safe drivers?’ and reviewed insurance data that demonstrates major crashes have causality with minor crashes. It continued with a description of Heinrich’s Triangle of Occupational Safety and a comparison to on-road accidents. With these points as the foundation, Bin-Nun proposes a manner of measuring “near misses”, heavily applied brakes and similar events that could be used to predict the safety of a self-driving system well before the real word data of actual incidents is returned from the field.
STANDARDS AND REGULATIONS
It wasn’t just the panel I moderated, the issue of government involvement was a primary discussion on many presenter’s minds. While the government currently is allowing the industry to define its own path, Rosekind’s previous role at NHTSA made him uniquely qualified to comment on the government’s role. He identified a typical regulatory process as a minimum ten year time frame. Allowing the industry to take the lead on Automatic Emergency Braking, for example, is bringing that technology to consumers 4-5 years faster than if the government takes the lead. The panel discussion with Stockwell and Kaufman addressed this issue as well as what eventual regulations may look like, when the federal government will eventually take control, the differences between freight and consumer self-driving vehicles and the role of 3rd party testing houses in their certification. Unfortunately as moderator, I could not take notes, so in the interest of not relying on my memory and misquoting the panelists, I can only provide a general overview of their responses. Even though the two come from different aspects of the autonomous industry, they agreed on most things – the process will be slow, limited in area and scope, with the process being led by the industry for the near term.
Events like this are critical for learning about the state of technology and the open issues facing the teams developing them. It also supports the exposure to new ideas and with the amount of networking that happens, certainly new concepts are created after the event by the relationships created from the conference. AV19 helped me finally put all of my conflicting thoughts and ideas in the proper context. Here, in no particular order, seems to be the consensus about the state of the technology and industry. Level 3 and Level 4 vehicles are coming soon, but in very small quantities and in very limited geo-fenced areas. Level 5 is the ultimate goal of everyone, but the technology is not there yet and there is no real agreement on when it will happen. With all these things in mind, there are two concerns I cannot shake.
ISSUE 1 – PASSSENGER/REAR SEAT SAFETY
The crash-worthiness of vehicles is currently measured on driver and front seat passenger safety, the area of back seat safety is secondary. Level 5 vehicles basically make everyone a passenger, and with that in mind there is likely a revised evaluation of passenger safety coming with commensurate new standards and regulations. Having driven a RV over many many miles, the similarities with Level 5 autonomous are striking. Like the ability to move around the cabin during transit that will come with the extra space that comes from eliminating passenger controls. And the likelihood of new seating positions, like sideways seats or even potentially laying down. Self driving cars will not eliminate all accidents, so consideration about these new passenger environments needs to be deeply thought through.
ISSUE 2 – HUMAN DRIVER SKILLS
If passenger safety is a critical component of the new world order, driver education is the serious missing piece of the puzzle that keeps me up at night. Let me explain. Level 5 (fully autonomous) is years away while Level 3 (autonomous only under the exact right conditions) is imminent. Right now Level 3/4/5 self driving systems struggle with certain aspects of daily driving; Merging into rush hour traffic is difficult, reacting in time to last minute emergency maneuvers from other vehicles is problem-some (following a car at highway speed that swerves last minute for road debris for examples), or driving in snowy/icy road conditions (2-4% of accidents and deaths). If you have ever been in the back seat of a self driving car when it hands control back to an inexperienced driver, you totally understand where my concerns lay. (full disclosure – I was so uncomfortable I asked to get out).
Add to these aspects are the fact that sensors do not yet have the range necessary for software to calculate all possible scenarios during high speed driving. In the near term where Level 3/4 systems become commonplace, the vehicle will always throw the control of a vehicle back to the driver to manage those situations where it is incapable of making a proper decision. Look at this video of Level 2 Tesla Adaptive Cruise accident to understand the issue that bothers me most.
Two things I know through my year’s of driver and safety instructing is: a) young drivers make decision errors because of their brains and b) the more someone drives, the lower their accident rate and the safer they become.
What will happen to the accident rate of humans if the computer manages to drive them around during the 90% of the easy circumstances a vehicle encounters, and only hands control over when the situation is extremely difficult to manage? With this as the new norm, a 40 year old driver may be asked to drive in snowy conditions when the environment becomes too difficult for the software and sensors. If the previous data holds true, the accident rate of these drivers will actually go up on a per miles driven measurement. So the end result is that all these advanced systems will actually increase the accident rate of humans while at the same time reduce the number of accidents that happen during normal situations. And while I have no doubt the total death rate will drop, consumer sentiment is a little more unpredictable.
Full level 5 autonomy has to be the goal and objective. But in my opinion the interim steps with Level 3 and 4 will cause many issues that we as a society have not fully evaluated and accepted i.e. an indirect but measurable increase in the human accident rate per mile. The marketing hype that permeates the technology is disguising some realities; accidents will still happen, people will still die, and the question of who is to blame will still be open for interpretation. These issues, and more, could also be the reason some manufacturers are working towards Level 5 and deciding to skip the interim steps. One thing is sure, conferences like AV19 are a necessary part of the resolution of these issues. If you want to learn more, we ask you to join us at AV19 Detroit this August.
PS Many thanks to Joel Adams, Founding Partner, Hacker Studios and President, VENDEV Management for his event pictures. His pictures turned out much better than mine and really bring this article to life.
And here are a few more pictures courtesy of the IQPC Team