I got a message on my phone from an unknown number. In a soft alto female voice the message said:
We’re sorry, an application error has occurred. Goodbye.
Earlier this month I mentioned that the streets of San Jose are awash with robot cars (five out of a sample of several thousand along a particular high-robot-volume street qualifies as “awash”).
I mentioned that one of the cars was a test vehicle for an outfit called Nuro, and I further mentioned that almost all the content on their Web site is a big treatise on safety. I went on to say that I had not read that document.
Well, today I was thinking more about it, and I went back to Nuro’s site to poke into their safety information. First impression: A document for non-experts that tackles very complex technical issues, but it seemed pretty legit.
Final impression: Nuro must have some pretty serious cash behind it, to take this long and winding road to achieve public trust.
The safety paper opens with the observation that 20% of car trips in the United States are people going shopping, and another 20% are people running errands. For many of those trips, the human is there simply to ferry stuff around. If robots can accomplish that task, that directly reduces the exposure of people to injury in automobile accidents — they’re not in their cars at all! Instead they are home moving the American economy forward by playing Candy Crush.
Nuro also mentions near the start of the document that 94% of all traffic accidents are due to human error. Remember that number when someone someday says, “30% of all robot-car crashes are due to software failure!
Nuro is creating a vehicle that doesn’t have people inside it. That gives it some very interesting advantages in the safety realm — the vehicle can choose to crash into a light post rather than hit the idiot that ran out in front of it. Self-sacrifice is an option for a vehicle without people in it. And the vehicle itself can be squishy, since it doesn’t have to protect occupants. The “windscreen” is a shiny panel on the front of the vehicle designed to give humans visual cues about the behavior of the car, but it doesn’t have to be layered tempered glass. It’s just shiny bouncy plastic.
Not having an impatient human to appease means the robot can putter along at a speed that increases decision time and shortens stopping distance. I think that’s important… but 25 mph max might be a little too slow for the streets around here, until we can get rid of all the impatient humans.
There are many, many words used in the document about when the robot decides it can’t operate safely and will pull out of traffic until a remote human operator can take over. While I see the necessity of that short-term, I expect with a few improvements to civil traffic control (flagman signs that can interact directly with robot cars springs to mind), that before too long the robots will learn to outperform the human backup.
I chose the word “learn” because there is a sort of cyber-attack I had not heard of before. You have probably heard of machine learning, although it’s frequently (and incorrectly) labeled artificial intelligence. Many companies have developed sophisticated systems that, after exposure to countless examples, are able to generalize information. It’s super-slick.
Nuro’s cars work that way. They are constantly gathering data from the environment and using that to refine their behaviors, and they share that information with the rest of the fleet.
But when your data comes from the environment around you, assholes can manipulate that environment to teach the machines falsehoods. Sometimes yield signs are octagonal and red, things like that. (Although to be successful the false data would have to be about something subtler, I suspect. I can easily imagine college-me arranging traffic cones differently every time a Nuro vehicle passed by. It’s an obvious parallel to my “yeeech” experiment, which shall not be documented in this episode.)
Of course there’s all the other usual stuff to keep the vehicles from being hacked, and one advantage of “safety as a priority before the first line of code is written” is that security also can be built in at the ground level.
Also mentioned more than once: the “whole widget” concept. If the software and the hardware are developed together for a single focussed purpose, it will work better and be safer. Steve Jobs would be proud.
And if you consider air quality to be a safety concern, then something like this makes everyone safer.
Nuro recognizes that the biggest obstacle to their success is social. Will people seeing Nuro’s placid robot cars poking along through the neighborhood think good thoughts or bad thoughts? Will appreciation of reduced traffic congestion, better air quality, and a more convenient life outweigh the fear of a robot uprising, and perhaps even worse, the fuming rage of being stuck behind a little robot car doing 25 in a 35 zone?
Today as I used public transportation to go to work, I saw five robotic cars, operated by three different companies.
Three of the five cars were Urban Automated Driving vehicles operated by Bosh and Daimler, running (human monitored) robot taxi service along the same corridor my bus takes. The Mercedes c-class vehicles are equipped head-to-toe in lidar units (lidar is like radar… with lasers!) and if I loaded the app I could ride in one. Which… is tempting, for purely journalistic reasons. My biggest question: How bored is the human monitor? Super-bored means things are going smoothly; super-bored also means that the human will never spot the emergency in time.
The second company was Nuro. The vehicle was a Toyota or whatever with sensors all over it, but what the company is actually developing is an autonomous vehicle that doesn’t have seats in it at all. Their dream: order your groceries and have the robot bring them to you. The vehicles are electric and since there is no need to account for human comfort, they could theoretically be much, much cheaper. It is easy to imagine that many companies that sell stuff would be interested in having something like that. Nuro’s Web site doesn’t have a lot of information, except for a pdf with a major discussion of safety (that I didn’t read).
There was a third, but my most humble apologies, dear readers, I don’t remember the company name painted on the car. It was not Google; I haven’t seen one of those in a while. Apple, should they even still have experimental cars, would keep them anonymous (which, as I think about it, would be just as definitive as putting a neon logo on the side — no other company would operate vehicles with a bunch of extra gear strapped on without missing the chance to brag about it).
As cities go, San Jose and the rest of the unplanned, disorganized sprawl that is Silicon Valley is… meh. And the cost to live in meh is staggering. But one thing I do enjoy is that it feels like we are just a little bit closer to the future here. And there’s nothing like Bay Area traffic to make you really, really, look forward to the day when people are not in control of giant deadly machines.