Hacker News Reader

Comment by OliverJones | Hacker News Reader

The CEO of Ford Motor Company, Jim Hackett, has claimed this: for autonomous vehicles to work resiliently at scale they require

1. vehicle-centered sensors

2. a way for vehicles to share their sensors with nearby vehicles,

3. detailed models of the cities in which they operate, and

4. wireless networks capable of transporting all this data to where it's needed.

He calls all this the "mobility operating system."

In the interview I heard (I wish I could remember where, sorry) Mr. Hackett said he believes the defensible intellectual property for autonomous cars will be those detailed city models.

So, I am delighted to hear that municipalities are getting ahead of this issue with respect to city models.

View on HN

Isn’t an issue with no. 2 that bad actors could spoof or inject bad data and cause problems for cars reading this data in good faith. You can discard one abnormal data point but not if it’s coordinated.

Do you worry about some random asshole kid hurling a re-bar in to rush hour traffic? REing automotive tech, setting up transmitters, and spoofing traffic data to throw a virtual re-bar in to rush hour traffic could become a pandemic unless we figure out a way to fight such attacks.

The difference between these two scenarios is that the asshole kid throwing rebar is easily detectable and to do it has to be at the location of the crime.

The asshole spoofing traffic data, depending on how the transmitters are set up could potentially do it from a long distance away, and may actually be hard to detect.

Emergency vehicles in some cities can emit a signal that changes lights from red to green. It's a felony to spoof those devices, but people still get caught doing it.

I'm sure a system can be made resilient to a single bad actor, but, yes, those attack vectors will have to be thought through.

relevant: https://xkcd.com/1958/

Forgetting to use your turn signal, or leaving it on, or running at night without headlights, is "injecting bad data" into the present system.

Yes. But that’s why you need the model and the car-centric sensors. “Trust but verify.”

If they need to verify, why do they need to talk to peer-vehicle sensors? Ideally what would happen is a car a quarter mile ahead gives the cars behind a heads up that something is different (construction, ladder or junk, stalled car ahead, icy conditions, etc.) but if this look ahead dats isn’t reliable of at least can’t be trusted, what are you verifying it against? Not your own sensors cuz then why even get this fleet data to begin with?

It can at least bias your priors ahead of time. When you (as a human) are using Waze or Google maps and you see a speed trap ahead of you, or a traffic jam, or construction or whatever, obviously you can't blindly trust that. But you can slow down and prepare, and if and when that occurrence comes to pass, you're better prepared to deal with it.

Trust but verify

Thanks. I was on the phone and autocorrect got me.

Are those not two different kinds of models though? For self-driving cars, the most important city models and those to which the Ford CEO was referring would surely be high-precision records of the physical layout of the city. Whereas the models in the article are simulating how people / cars / ... move around the city.

Hackett said the city models were likely to be slowly-changing stuff.

Human drivers also gather data from nearby vehicles. That's the point of brake lights, turn signals, and rearview mirrors, for example. And, yes, human drivers also spoof that data: mostly by forgetting to use those signals correctly.

The ideal is a high precision representation and then let that each model use what it requires. Better aggregated than extrapolated. This is the central idea of our modeling platform at Aimsun.

Maybe this post? https://www.vox.com/2018/1/9/16865384/smart-vehicles-cars-fo...

I just read point 3 and had an awful moment of thinking, "sorry I can't drive in Toronto. I didn't get the DLC subscription for its map data and I never learned to drive."

It's very strange that we're able to drive without 2, 3, and 4.

We can't drive without 2, 3, and 4. We already have them.

2. a way for vehicles to share their sensors with nearby vehicles,

The sensors are the driver's senses. Turn signals, brake lights, and horns relay the consequences of those sensors to other drivers.

3. detailed models of the cities in which they operate, and

Most drivers in any given place have a detailed mental model of the streets they use. Signage contributes a lot to the model. Traffic lights contribute temporal information, as do orange barrels and other ways of noting temporary alterations to the model.

And, consider London taxi drivers. They have the "knowledge," a detailed model of their city.

4. wireless networks capable of transporting all this data to where it's needed.

Sight and sound are generally reliable transport-layer network media.

Reply on news.ycombinator.com