Hacker News Reader

>It almost sounds like the system will brake at newly installed traffic signs.

One of the only situations where both "brake" and "break" have the same meaning in a sentence.

Or, worse yet, will crash into an actual obstacle where white listed sign is.

Self-driving cars are entirely reliant upon mapping data for all sorts of functionality. For example, how is it supposed to differentiate between curved roads and on/off ramps? Decent navigational maps require huge amounts of manual intervention, which is why Apple's mapping software initially sucked. The maps for self-driving cars will require an order-of-magnitude more data.

Yes, but Apple and Google, they only have what their company cars bring in for road data. Tesla makes it clear every customer is also acquiring data for all their cars to make better decisions:

> Curve speed adaptation now uses fleet-learned roadway curvature

So they have an advantage over Google and Apple in that they have mass market cars full of sensors being driven for them for free.

Apple and Google have billions of smartphones collecting location data. Not all of the data are currently sent back, but they can certainly do it with just a software update. For example on my iPhone under location services it already lists "routing & traffic" as one of the system services using location. It's easy enough to extend it to mapping too.

Unfortunately that's only location data. What cars with more sensors can send is closer to: this location has N lanes, turning radius is M, the slope is X, I'm travelling at Y mph, I see connected road at location Z, I see the following signs...

What your phone can send is only "I'm more or less at X, I'm moving fast so we may be in a car. But I may as well be in a glider. Who knows..."

I guess so... Google's mapping cars have expensive sensors that won't make it into production cars. And if they partner with a single auto manufacturer or buy up Lyft, they will have just as much data as Tesla has.

Yes, this seems really complex to me, too. I'm also not quite sure on why exactly it is so complex to distinguish between objects (including vehicles) on the road and ones above/next to it. If they have radar images (I imagine them as images with depth information, which might be fundamentally wrong), they should be able to tell both where the road is going and, with that information, which of these objects are of relevance.

But for the learning part, they probably combine the camera that they used to date as their primary device in combination with the radar (at least in daylight scenarios) to identify objects. They may even be able to learn about the special material properties, like the reflective coating of a traffic sign.

> I'm also not quite sure on why exactly it is so complex to distinguish between objects (including vehicles) on the road and ones above/next to it.

Because radar does not have the same resolution as LIDAR.

EDIT: Phased array radar and cheap stationary LIDAR should get under $100 in ~5 years, at which point this whole argument will be moot. Hacks in the meantime!

What does the world look like in radar? Are there any visualizations available?

Could multiple radar emitters and receivers be used to create a phased array and improve resolution?

Here is the resolution calculation for aircraft radar (with respective example of signature - aka image )

http://www.radartutorial.eu/01.basics/Range%20Resolution.en....

For angular resolution: http://www.radartutorial.eu/01.basics/Angular%20Resolution.e...

> What does the world look like in radar?

Depends on the radar system. Some are distance only without direction. Some are 1D (a line, usually horizontal) and some are 2D. Many objects are partially opaque, which is confusing. Resolution is very poor compared an optical device of the same size.

> Could multiple radar emitters and receivers be used to create a phased array and improve resolution?

Yes. However, this is currently bulky and expensive (in dollars and in compute power). Thankfully, it looks like capitalism is coming in to the rescue here and miniaturizing the everloving shit out of complex radar arrays for human interface tech. This should be usable for vehicles as well.

All the automotive radars are phased array devices, and have been since the Eaton VORAD of the late 1990s. No moving parts. They're usually 1D scan (horizontal) only, although 2-axis scanned automotive devices exist.

If you're cresting a hill or rounding a curve, passing under a bridge, etc., static objects in the visual field are translating in front of the vehicle and are not easy to distinguish from slow moving vehicles. With simple sensors, there isn't enough information to decide that these blips are on a collision course or not.

The Florida situation was the easiest set of circumstances for machine vision to handle and it still failed. More complex, dense road systems with real terrain are much harder to handle. This radar system will still get someone killed.

The driver is still the final authority. Pay attention and in most cases, if you die, it won't be your fault.

Is LiDAR worthwhile despite rain and snow detection? Can that be ignored in software?

I'm also not quite sure on why exactly it is so complex to distinguish between objects (including vehicles) on the road and ones above/next to it.

In the case they're talking about here, it's because to do so you need to predict where the non-visible road surface is going to be.

Consider travelling up a continous, slight incline. Precisely at the crest of the incline is an overhead gantry sign, positioned such that for an observer travelling up the incline it is located directly in line with their current direction of travel.

The only way that you "know" that the road actually dips under the sign rather than the sign being on the road surface is experience.

You can also see other cars "disappearing" under the sign. And the sign going up a bit as you move closer.

That's the kind of higher level reasoning that's easy for people and hard for machines.

  Upon closer reading, he explains it somewhat.

The author of the article is a nameless collective, not an individual. They explain it, but there is no "he" to ascribe the explanation to. We do not know the names of the authors.

Annoying you are being downvoted. You are entirely correct, and it is worthwhile to fight against the nonsensical notion that Elon === Tesla.

For what it's worth, there is some strong evidence that Elon penned this post himself - not that it is clear from the post itself.

https://twitter.com/elonmusk/status/771446048262946816

> Finishing Autopilot blog postponed to end of weekend

https://twitter.com/elonmusk/status/774155658476212224

> Will get back to Autopilot update blog tomorrow.

https://twitter.com/elonmusk/status/774664927835553792

> Will do some press Q&A on Autopilot post at 11am PDT tmrw and then publish at noon. Sorry about delay. Unusually difficult couple of weeks.

I was already aware of these tweets, but that doesn’t mean he wrote it all himself, just that he is involved in it. The fact is, the blog post has a byline and the byline does not mention Elon Musk or any individual person's name.

I got this from twitter, where Elon Musk writes stuff like "Writing post now with details." and "Will get back to Autopilot update blog tomorrow." So I assumed he's authored the thing.

Yup. As soon as you're white/blacklisting you've lost at AI because you're assuming the rules never change, and if that's the case then why use a learning AI in the first place.

I wonder how many fatalities it'll take before they realise that 'the driver should've been paying attention' isn't a good enough excuse.

Incremental steps. They will move past whitelisting in time.

Note that humans use whitelisting too, to the point that they do not notice when a stop sign has been defaced or converted to a "give way".

I think you have it backwards. The radar will not initiate braking events for newly discovered stationary objects. Once it has human feedback from multiple sources, the object will be whitelisted and cause braking events.

blacklisting would cause the behaviour you describe.

I read it the same way as karyon. Whitelisting something means recognising it in a known-good way (in this case, "we know that's an overhead sign"), whereas blacklisting means known-bad.

The blog post says "When the car is approaching an overhead highway road sign positioned on a rise in the road or a bridge where the road dips underneath, this often looks like a collision course. [...] If several cars drive safely past a given radar object, whether Autopilot is turned on or off, then that object is added to the geocoded whitelist."

So as i read it, the system does think it will collide but ignore it if the whitelist says the object is safe.

The paragraph after that says basically once they have enough data, the system will start braking on unknown objects with gradually increasing force as the confidence level rises. So basically, it will brake on unknown traffic signs but only slightly, as the confidence level shouldn't get too high.

Basic problem: the system doesn't know where the road surface is. Google's vehicles, and most of the DARPA Grand Challenge off-road vehicles, profiled terrain with a LIDAR. They know where the road surface is. This is essential to off-road operation, but on-road, optimistic assumptions (no big potholes, road not leading off cliff) can be used, at some risk.

What about a tractor trailer (like Florida) in the same spot as a white listed traffic sign?

Road signs will be more to the left or right, truck will be in front.

Corner case: truck jackknifed across the road immediately under a white listed gantry.

I read this as saying: we start braking lightly when detecting an unknown object, applying more braking force as it becomes clearer that this object is a collision risk.

Once enough human or autopilot trips detect the same signal with no collision, whitelist the signal and cease braking-and-monitoring behaviour.

I am probably misinterpreting but that is what makes sense to me.