Self Driving Cars?
By Robert Bernstein
Self Driving Cars? Horus Muchen Wu at Humanist Society of Santa Barbara - 1-18-20
Horus Muchen Wu is a senior software engineer working on one of the cutting-edge areas of research in Silicon Valley and around the world: Autonomous Vehicles, also known as Self-Driving Cars. Horus asked that we use the pronouns they, their, them so that is what I have used from here on.
Horus took time out of their busy schedule to fly down and speak to us at the Humanist Society of Santa Barbara about the technology and about the social issues of Self-Driving Cars.
They made it clear at the start that they were expressing their opinions only and not those of their employers. (Horus is the only person I know who left Google for another opportunity, by the way.)
Why talk about Self-Driving Cars? The industry is not making money; in fact, a lot of money has been invested with no return so far.
Horus started with the issue of CDC data showing that unintentional injuries are the #3 cause of death in the US. Car crashes are like a 737 crash 5 days/week every week. Few people would take a 737 at that risk level!
Then there is the matter of time wasted in traffic. In the San Francisco-Oakland area 103 hours per year per auto commuter are estimated to be wasted per in traffic. This, according to a Texas A&M Urban Mobility Information Report: https://mobility.tamu.edu/umr/
162 lifetimes are wasted each day in traffic in the US. I am surprised the numbers are not higher.
Horus now walks to work. Their boyfriend commutes 1.5 hours/day each way. Caltrain takes 45-57 minutes between Mountain View and San Francisco for the Baby Bullet. And it is rare to get a seat. And it is expensive. If that Baby Bullet is missed, the regular train takes 75 minutes. Horus gets carsick if someone else is driving.
Horus referenced a book called "Being Mortal" about end of life issues. The mother of Horus still lives in China. The mother is one of four sisters. When the mother's mother was hit by a car, there were were four daughters to look after her. Horus was born under the One Child policy. The mother and father are divorced, so Horus has to take sole responsibility for each of them. From an ocean away.
The mother gave up driving. The father still drives. In the US you are isolated if you can't drive and this accelerates aging. China has better public transit than the US. But the transit infrastructure in China is less friendly to the blind or those in wheelchairs than in the US. Driving is a high-risk activity. Just one accident causes doubts about abilities.
Why fully self-driving cars now? Why not take an incremental approach? Cruise control has been around for many years. Adaptive cruise control allows this to work in traffic, keeping a safe following distance. Lane control is another feature becoming more common. But the latter can lead to mistakes, depending on the strategy used. With some strategies, a car on the freeway may follow an exiting car ahead off the freeway instead of continuing straight ahead.
Horus laid out five stages of automation:
The most dangerous of the five is the middle one: Level 3: Conditional Automation. The driver must be ready to take control of the vehicle at all times with notice. Most of the time the car seems to be safely in control. People become tempted to put on makeup or reach into the back seat to get something. Horus showed a video of a couple sleeping in their Tesla speeding on Highway 90 near Boston. With the car in control it is hard to stay awake and attentive.
Safety requires either lower levels of automation or a leapfrog to Level 4: High Automation. The vehicle is capable of performing all driving functions under certain conditions. How hard is it to get to this level?
Horus showed a video of the 2005 DARPA Grand Challenge. Autonomous Vehicles had to drive from Los Angeles to Las Vegas on a dirt road. It seemed that the challenge had been met and we were close to success. The vehicles could handle a winding road with sheer drops on the side. A human would get flustered by this, but the autonomous vehicles could just focus on what mattered.
But the problem in traffic is way harder. Horus showed a video of driving in San Francisco. Double parked vehicles. Delivery vehicles. Scooters, bikes, pedestrians come out of nowhere. How to decide how to get around stopped vehicles.
A reporter followed a Waymo autonomous vehicle in Chandler, Arizona, near Phoenix. Lane changes are agonizing to watch. Often the driver takes over. Waymo doesn't like its vehicles to be followed. In one case the human Waymo driver led the reporter to the police where the reporter was questioned and released.
How do vehicles communicate? Humans give gestures and make eye contact. One could imagine a kind of "chicken dance". What about an electronic sign board? This was banned by law as a possible distraction!
Unprotected left turns are the hardest. There are so many unpredictable things that can happen. The videos following the Waymo vehicle always seemed to involve a driver taking over. Chandler is way easier than San Francisco and even Chandler has not been mastered.
What is needed? Perception. Prediction. Planning. Control.
Control is relatively well studied in the field of robotics. Perception is a big challenge all by itself. Also known as computer vision.
Horus showed how the system has to put a box on everything of interest: Pedestrians, cars, speed limit signs. What about trees? A "semantic map" of the world is needed. Separate out things that don't change. Houses, trees. Pedestrians are marked green. Cars are marked red.
First it is important to map the city and the roads of interest.
But in real life you may not see a whole person. And what about a person on a bicycle?
Horus displayed the details of an actual accident that took place between a Waymo vehicle and a bus on El Camino Real in Mountain View. Waymo wanted to make a right turn and predicted the bus wouldn't try to pass it. But the bus did pass it and Waymo moved left just at that moment. It is not clear why. No one was hurt, but it was a collision.
An Uber car did cause a fatal accident. It misidentified an object, yet it kept on going. It was a woman on a bike.
Tracking is a key issue. If two humans are walking together it is important to keep track of their positions. Seeing half a person does not change the fact that it is really a whole person walking with another person. The Uber car didn't do that well. Waymo cars do better at waiting until they understand the situation.
When will we get to Level 4? We hear both sides. Some say this year. Some say not in 30 years. A lot is perception. The current systems are based on "deep learning" which is very data greedy. Humans learn the basics of driving in 100 hours of practice. A child can see one cat and not only recognize other cats, but we can recognize that a tiger is a cat. Deep learning may need to see tens of thousands of cats to recognize cats. This is not practical for autonomous vehicles.
We don't really understand how we do many of the things we do with ease. We walk on two feet. The Internet is full of videos of biped robots falling in embarrassing ways.
But even before having fully autonomous driving, there are things to be done. Driving trucks on a highway has fewer surprises. Tu-Simple has a video of their attempt at this. USPS has a contract for this. It is easy for robots to beat humans at this. Humans fall asleep and lose attention.
Buses are also easier than cars in some ways. Horus showed an ST Engineering video of a proposed mini bus. It runs a fixed route. It can have dedicated lanes with signal communication. A bus can have many sensors so there is no blind spot.
Horus also played a video of an Amazon Scout small delivery vehicle that would run slowly on sidewalks. It has low risk of injuring anyone. It could provide economic value by eliminating the need for humans to walk that last bit, going door to door.
What happens if we succeed? Horus showed this map of the top job in each US state. In a majority of states "Truck Driver" is the most common job.
Horus also showed the hourly cost of parking in major US cities. New York was top at $27/hour. Autonomous vehicles would eliminate the need for parking as they would keep circulating. The cost per mile of a shared autonomous vehicle is just 17c/mile compared to 57c/mile for a personal vehicle according to one estimate.
Horus then answered some questions. What if an animal runs in front of the car? The car has sensors all around. It can see what you cannot and it will stop.
Ron asked about having a smart road? China is working on that. It costs money and is not always the best use of public money.
Is the bus video real? It is conceptual, but quite far along. Still, a lot of unexpected things can happen even in a designated bus lane.
We have self-driving planes? Yes, but that is much easier in steady flight. And we never got rid of the pilot.
Weather? All of the tests have been in nice weather in places that have nice weather most of the time. "Snow will be a challenge." But Horus thinks it will be easier to add weather accommodation once good weather driving is mastered.
Wayne asked about the unfair higher standards for autonomous vehicles than for humans. Horus agreed it is not fair. People worry about their flight on their drive over to the airport. When the drive is usually more risky than the flight!
Mayor Sheila Lodge asked how this would help the toll plaza bottleneck on the Oakland Bay Bridge? Autonomous vehicles could bypass the toll plazas completely with a transponder.
What about the Boeing Max example? A good question. Many tech nerds don't spend enough time learning ethics or management. Different companies are approaching this in different ways.
I will add that several of us were privileged to have lunch with Horus before the talk. We learned that Horus wanted to use the pronouns they, their, them for a reason that is not often discussed. Horus said that in job interviews there can be gender bias and these pronouns help reduce that bias.
Horus originally intended to be a tech manager and was writing software both a hobby and as a step on that path. Horus thought they could hardly be outstanding as a software engineer, but could more likely be an outstanding manager and make a bigger impact. After a few years of practice, it turned out that they actually developed impressive software skills, and had a good shot making a difference being a software engineer instead of a manager.
I should add that this issue of autonomous vehicles is of great interest to me. Sustainable transportation has been of interest to me for a wide range of reasons. In the 1970s as a child I saw the inevitability of running out of oil and other fuels. In the early 1980s I became aware of the climate crisis and the urgent need to stop using carbon fuels even if we were not running out.
I then came to understand the social harm caused by heavily subsidized forced dependence on automobiles in the US. Even if they ran on a free, unlimited, non-polluting energy source. One third of Americans cannot drive and they are severely limited as a result.
I wrote a chapter called "The Speed Trap" in the popular Shorter Work Time book called "Take Back Your Time". I have made my chapter available here: http://swt.org/oshare/
In my view autonomous vehicles could be good or bad, depending on the economic fairness of the incentives. The advantages would include: Those who cannot drive would have affordable door to door transportation. Half of urban land in the US is paved over for motor vehicles or parking. Shared autonomous vehicles would reduce the need for parking. Users would pay by usage which tends to reduce usage. Shared autonomous vehicles also separate out the attachment people have to the status-symbolism of car ownership.
I would rather see investment in good public transit. But autonomous vehicles could still solve the "last mile" problem of public transit, making transit more efficient and practical.
Another possibility is that autonomous vehicles can form "trains" on the highway which would reduce air resistance. At highway speeds, most of the energy is used to move air. Closely following vehicles can minimize this load.
My biggest concern on the down side is that autonomous vehicles could encourage more rural and suburban sprawl. That would be the worst possible environmental outcome. It all comes down to charging the true cost for environmental impacts.
A government report (Office of Technology Assessment; Saving Energy in U.S. Transportation; U.S. Congress, OTA-ETI-589, 1994 page 118) showed that we subsidize private motor vehicle use in the US at a cost of about a trillion dollars each year in various ways.
Here is that report: https://www.princeton.edu/~
If more Americans were aware of that staggering number, our conversations about land use, transportation and public policy might be very different.
Here you can see what else the Humanist Society is doing and what it has planned for the future. Everyone is welcome!