Comment by phkahler | Hacker News Reader

Comment by phkahler | Hacker News Reader

Parent Full thread

>> I really want to see fully autonomous vehicles but I don't really believe we'll see them on the road in the next 50 years.

I agree with you 100 percent. DAS (driver assist systems) are getting to be very common. I've worked at a couple companies that produce them, though I did not work on those products. I rode on the highway in a car outfitted with prototype lane keeping, and it was not quite as steady as I'd hoped. They also had an option that would apply a light force to keep you in the lane but could be easily overcome (and I think went away if you used a turn signal). It felt like a sort of speed bump between lanes. It wasn't trying to be a really complex system, and probably has some real world value - in reasonable situations I'd like to be able to use both hands to eat some food from the drive through, while keeping my eyes on things.

I've also seen video (circa 2004) of an autonomous car driving 100kph down a winding dirt road, staying on its side of the road and automatically stopping. But again, the engineers wanted to do so much more with it, but the rational guys in safety would not allow it.

I also competed in the AUVS autonomous ground vehicle competition in 1994. I wrote a lane follower by taking 20-30 lines of pixels off an NTSC frame grabber on an Intel 486. The core algorithm was on the order of 100 LoC, and we took second place. Super simple algorithm, hardly intelligent.

There is a huge difference between a PID controller maintaining position in the lane with radar assisted cruise control, and being a fully autonomous vehicle fit of unaided driving on public roads. There is a whole range of system capabilities, and the public has no idea what's in any given car. Comparing highway fatalities/mile to national statistics covering all roads and conditions is bullshit and a certain company that's recently killed some people knows it.

But hey, we all want to be able to read a book on the road or have truckers take a nap on long hauls, so let's keep deluding ourselves that this stuff will be ready for prime time in the next couple years.

"I wrote a lane follower by taking 20-30 lines of pixels off an NTSC frame grabber on an Intel 486. The core algorithm was on the order of 100 LoC, and we took second place. Super simple algorithm, hardly intelligent."

How is this relevant to self driving cars in 2016 and the following decades?

I think he/she is arguing that what people are "ooing" and "ahhing" about is not really that amazing or novel, it's the infinite number of edge cases that will kill people regularly that is difficult and will not be solved for quite some time.

That's why I disagree with Elon Musk and others who say "autonomous driving systems only need to be like 50% or 2x better than the average driver".

What if I drive 5x better and safer than the "average driver"? Doesn't that mean that these cars would now put me in more danger than I would normally be in with my own manual car?

Self-driving cars should be 10x better than the best (as in safest by number of incidents, say, per decade) drivers on Earth. Then I could begin trusting them ... if only there wasn't that pesky security issue, too, with self-driving cars getting hacked and then driving you off the highway at 70mph and killing you.

We see Teslas getting remotely hacked. If that happens with Tesla, a Silicon Valley car company with a lot of hired top tech talent, how can we ever trust a company like GM or Volkswagen to build safe self-driving cars?

But even if you're the best driver around you 19/20 times you still drive now.

I don't see why you give wetware more leeway than hardware.

> What if I drive 5x better and safer than the "average driver"? Doesn't that mean that these cars would now put me in more danger than I would normally be in with my own manual car?

Because you are unlikely always that.

Are you tired? Do you have a cold? Are you hungry? Are you playing with the radio? Are you having a conversation with a passenger?

The huge advantage that self-driving cars will have is integrated systemic control from multiple sensor sources and multiple transducer outputs--and they will never get distracted.

Human drivers have mostly 1 input: vision. Human drivers have mostly two outputs: hand and foot.

The lag in that system from input to output is >100ms even if you are perfect (I will grant you an exception if your name is Michael Schumacher).

Automated systems will have far more inputs. Multiple vision sources. Multiple radars. Temperature readings. RPM readings. etc.

Automated systems have far more outputs. Steering is the same. Brakes on individual tires is the big starter. With electric, individual suspension becomes feasible. Transmissions change form and become electrically controlled.

And the lag in the system is probably <10ms. That's huge. That's the difference in reacting in roughly 3 meters (2.7 meters) vs 1/3 of a meter at 100kph (9 feet vs 1 foot roughly for 60mph).

Finally, automated systems will take subtle preventative actions all the time. They will "see" the child on bicycle on the sidewalk (the idiot detector as Google called it) and "know" that they need to find him again when they attempt to turn right or slow down significantly. They will see the dude on a skateboard and know that they need to give him an extra couple feet if they can in order to deal with him falling. I know very few drivers that are that diligent.

yeh but humans have full AGI unless the culture comes calling (giz us a job dizzy) it isn't going to happen soon

AGIs aren't necessary to drive safely; a sufficiently advanced ANI will do.

In all conditions? not all driving is on a clear day on a motorway.

An AI that would be able to drive in all conditions (but not be able to do anything other than drive the car) would still be an ANI.

If you are 5x better than the average driver, the most likely cause of any accident you are in will be the other driver.

Why wouldn't you want to make them better?

Well ideally everything should be perfect but if a 2x system cuts US fatalities from 30k/yr to 15k are you going to ban it?

> What if I drive 5x better and safer than the "average driver"?

Then don't use the system?

You are giving thresholds - 50% to 2x better, 10x better. Before what? Before we ban manual driving? "10x better than the best driver" might make sense. Before we allow it on the road? Why not "better than the worst drivers we allow on the road"?

The most relevant point for Musk is probably "before the marketing works out". The market of people who consider themselves "substantially below average" drivers is much smaller than the market of people who consider themselves "not all that much better than average".

> What if I drive 5x better and safer than the "average driver"?

How do you know that you're actually 5x better than the average driver and not just deluding yourself?

Yes, that reminds me of the stretch on the 101 near Palo Alto that has two sets of lane markings (one old and slightly faded out but still very visible, and one newly repainted). Of course, drivers should only follow the newly painted set of lane markings.

Even as a human, I get confused while driving on this stretch. The old lane markings aren't quite faded out enough to easily tell them apart. As I'm driving, I see every other car slowing down and "dead reckoning," too.

How will the run-of-the-mill self-driving car's lane following handle this corner case? If it's a Canny edge detector (like most self-driving cars), it's definitely going to fail and find false positives.

Why do you assume most self-driving cars use Canny edge detection? And even if they did, you can simply have the algorithm choose the higher threshold (with some other sanity checks, suck as are the lines the expected distance apart) and ignore the lower threshold.

The algorithms for lane following are quite a bit more sophisticated than that though. Here's a paper from 2008, and there are other approaches as well.

http://www.vision.caltech.edu/malaa/publications/aly08realti...

Hough lines/RANSAC are definitely on the same level of ad-hoc sophistication as Canny edges (in fact, Canny edges are often done as a preprocessing step for the voting scheme used by the candidate Hough lines).

Take a look at Figure 12 for an exemplary look at its failure cases -- cases that are much less forgiving than the aforementioned Palo Alto stretch 101 lane markings.

I'd have thought road-marking interpretation would nowadays be based on something more elaborate, such as HOG, without the need for an initial edge-detection pass.

(But in this case, I'll add that HOG doesn't solve the problem of overlapping old/new road markings)

>> How is this relevant to self driving cars in 2016 and the following decades?

My point is that basic lane following can be done quite trivially, but it's nowhere near something one should entrust with the safety of human beings. Your question implies that you think all modern systems being deployed involve some really sophisticated AI. You don't know that.

As an example, anything using LIDAR is probably fundamentally incapable of robust object identification and tracking.

Is that statement backwards?

>> Is that statement backwards?

Nope. LIDAR is used to determine geometry when a camera system can't tell WTF it's looking at.

But you stated that anything using LIDAR was not capable of robust tracking. Perhaps you meant that anything using LIDAR was doing so because camera-based tracking is not very robust. But that's not the straightforward interpretation of your statement.

>> But that's not the straightforward interpretation of your statement.

Yes, that would be a communication problem of mine. It's something I'm working on ;-) But I would counter that LIDAR isn't going to fully make up for the problems of the camera based system with the result being a simple sounding statement like "if they're using LIDAR it's not good enough", or grossly oversimplified to "LIDAR sucks".

I will admit that LIDAR could conceivably enhance a really good vision system by offering better depth perception, but I don't believe that's why people are using it today.

Nothing in your comment makes me question the value of upcoming autonomous cars. It seems you're basing this on your experience with obsolete tech (aka the tech that currently on the road built by the non-tech companies). Even Tesla's assisted driving is being outmatched by a guy in his garage [1] doing full-data collection + deep learning. The more data collected in each ride the more the systems can improve their driving ability.

This stuff is limited right now but the sooner these (actually stage 4 autonomous) cars starts hitting the road and being used in real life, the faster the tech will improve. I expect some dramatic improvements once Google car quality sensors are combined with full-data collection on hundreds of thousands of daily rides.

Of course, it won't become totally autonomous until the vast majority of the cars on the road are using assisted driving and potentially interfacing/coordinating with each other. At that point the speeds can be increased and the driver can be disconnected from the backup driver role.

With kits already coming out with the idea of 'back-porting' old cars to use new automated driving tech, I don't think it's a huge barrier to adoption. It won't require everyone having to go out to buy new mid-level cars to participate. People can purchased insurance (and personal safety) incentivized kits to make their cars automated.

[1] Comma.ai https://www.youtube.com/watch?v=KTrgRYa2wbI and technical talk: https://www.youtube.com/watch?v=Hxoke1lDJ9w

>> aka the tech that currently on the road built by the non-tech companies

The level of arrogance in that is pretty staggering. Remember, I mentioned a system driving on (unmarked) dirt roads at highway speeds 10 years ago. I'm sorry, but advanced tech does exist outside of California.

> But advanced tech does exist outside of California.

And tech companies and AI experts also exist outside of California? What's your point?

I specifically addressed your pessimistic comment that dismissed upcoming driverless tech due to your experience testing the current DAS (driver assist systems) built by non-tech companies. I question the relevancy when tiny companies like Comma.ai have demonstrated that widely available deep learning algorithms allow a single guy in his garage to surpass Tesla's own tech. Therefore if that obsolete tech (which is itself is using 3yrs old sensors) is your point of reference plus some 10yr old videos... I'm not sure how relevant that is to what tech we can anticipate in the next 3yrs.

George Hotz' talk opens with the claim that "Tesla's and other companies current DAS tech sucks. They don't even capture the video/sensor data from the rides". Which I fully believe. It shows the wide disconnect between the limited DAS tech being rolled out by car manufacturers and the true capabilities of current technology from people working on real autonomous vehicles. Instead of glorified cruise control with lane switching.

The real question and risks seems to be around price and implementation details. Not whether or not the technology will actually work as intended within the next few yrs.

The system handling dirt roads was an order of magnitude more impressive than anything else I've seen including George Hotz little toy project. Can geohotz system navigate dirt roads? Intersections? Can it read signs? No, no, and no, but hey it can record video... I gave several examples of different levels of tech - all developed by non-"tech" companies.

I'll take the other side of that bet. My view is that it will come down to the insurance companies. They are going to price policies such that only a very wealthy person can consider self driving. This will be in less than 15 years.

Why would insurance cost more than it does today? Presumably new safety systems will trigger discounts to the degree that they cut the number of dollars that insurance companies need to pay out on their policies. But I see no reason that a profitable policy today wouldn't be profitable in the future just because other cars have more advanced safety systems.

Scarcity? If 99% of the cars are autonomous, and the only people left manually driving are driving Sports cars - everything from Mazda Miata's to Ferrari Enzo's, then that will be a very small, very expensive to market insure.

Don't forget 4x4 and vintage vehicles. The number of miles driven is one of the biggest predictors for claims and I imagine that most of these enthusiast manually driven vehicles will have low annual mileages. The insurance rates will almost certainly go up, but I do not think they will be unaffordable.

If 99% of cars on the road are autonomous, it will probably be very difficult to hit one with a self-driven car, so liability costs won't be high. And unless Miata parts and labor skyrocket in price for some reason, comprehensive and collision costs aren't likely to be high either, and one can always skip both.

Reply on news.ycombinator.com