New GM Cruise Self-Driving Video Shows More Mastery of SF Roads | Hacker News Reader

New GM Cruise Self-Driving Video Shows More Mastery of SF Roads

driverless.id

186 pointswonderhowto171 comments

That's pretty cool, it's great to see GM not sucking at bleeding edge software. They're getting their money's worth out of that billion dollar acquisition last March. I think Kyle Vogt is GM's MVP.

Looking at the 2016 California DMV disengagement reports, Cruise is at 181 disengages in 9,776 miles of testing in California for 2016. This can be compared crudely to Waymo's 124 disengages in 635,868 miles for 2016.

Waymo's operation is 2 orders of magnitude more advanced than Cruise, but in terms of video demos, we haven't seen much of the environments Waymo's cars are capable of navigating. Waymo has only revealed their cars driving on suburban roads, it seems only Cruise is tackling hectic downtown driving head-on.

GM, with their Lyft partnership and Orion assembly plant is, I think, in a great position to deploy early fast with 1st gen robotaxis.

I suspect the Waymo miles are not really comparable to the Cruise miles since Waymo seems to mainly test around Mountain View vs Cruise in SF.

I suspect Waymo doesn't plan to launch in a very large city.

It seems that it would still need to work in places like SF and Manhattan though unless they get drivers to sign a contract that they will only drive in certain areas.

If you've cracked automated driving, I'm confident that you can geo lock where the cars will do it.

The other reason is that the stats focus on disengagements as a result of technology failures or immediate safety concerns, and don't show the "routine" disengagements where the driver switches out of automatic mode for a period to avoid running the autopilot in situations it isn't test-ready for. Different manufacturers will likely have different standards for ruotine engagements/disengagements, and may also have different standards for what they consider a safety concern.

Waymo cars frequently drive the same routes several times a week. In fact if you go to certain intersections such as El Camino Real and San Antonio Rd, during the work week you only have to wait a few minutes to see one or more go by.

Do you (or anyone) ever see Waymo cars downtown?

In Mountain View (off Castro St), yep, all the time. In SF, no.

I think this is true of most of the self-driving cars. Certainly I've seen Uber self-driving cars doing the same routes over and over. And I think I see GM Cruise cars in the same part of town, but they may just be based nearby.

Do you think the routes are designed so that the car encounters specific obstacles the teams are working on, or do you think its more like they are doing extensive mapping of an area to make the autonomy a bit easier?

I suspect it is so that it is easier to measure the effects tweaks/development to the algorithms make.

> Cruise is at 181 disengages in 9,776 miles

> Waymo is 2 orders of magnitude more advanced than Cruise

Not to say that Waymo isn't ahead, but if you look at a graph of Cruise's disengagements vs miles over time, they experienced a rapid improvement. By November of 2016 they were at about 2.6/1000 miles. Better than the total would show, but still an order worse than Waymo though.

This article aggregates the miles/disengagement in a nice table at the bottom. Definitely worth a glance. Waymo is 8x better than the nearest competitor (which, as mentioned, is probably an unfair comparison since we don't know enough about comparibility of those miles driven) https://www.wired.com/2017/02/california-dmv-autonomous-car-...

A lot of the roads near me are two lane, but have cars parked on both sides. This requires a lot of hand gestures out of the window to other drivers as we negotiate the roads.

We also have a lot of single track roads with passing places. This also requires a lot of non-verbal communication between drivers.

Are there any videos of self driving cars attempting this sort of road?

Couldn't this be solved with an LED sign or something? Even if not, true self driving cars aren't coming for a while - they will still disengage and force the driver to take control in some situations. I'd be interested in seeing what happens in those situations, too - if you live near CCSF in San Francisco there are a bunch of these roads.

Why do you have to negotiate anything? If you're driving, you're higher priority than the car trying to leave its parking spot. It's their problem to squeeze into the traffic.

What do you do when two cars approach each other on a single lane road, with no room for passing?

I don't even know how a human would solve this. I would probably just avoid this road altogether.

Have you ever driven in the real world? Those roads are very common in many areas. Try driving in Malta sometime.

Yes. I just don't drive in areas like the one described.

I've never driven in Malta, but if I'm ever there, I'll be sure to try driving.

Uh, no. It happens all the time in Scotland, particularly the highlands. I ran into it in Iceland as well.

Generally, the car closest to the nearest spot suitable for passing just reverses to that spot. The other car passes, and they both go on their merry way. Very little in the way of communication required.

The person who didn't reverse is supposed to wave politely to the person who did reverse - that's the usual behaviour in Scotland.

Also its sometimes not just who is nearest who reverses back - sometimes things like corners, size of vehicle, if anyone has a trailer come into the decision making.

I assumed "don't be a jerk, wave to the other guy" and "don't make the lorry back up hill and around a curve" were given. :)

Of all the places I've driven, I've found Scotland (Perthshire and into the Highlands in particular) to be the most sane. Polite, little drama, and most follow the rules to a T.

I found it amazing how common this was in India (Kolkata) - and not only were they long roads, with cars parked on both sides of the road, their were some stretches that barely had enough room for one care to get through - as in, I couldn't open the door if the car stopped.

I have no idea how a Automated Vehicle would handle this - the humans had to engage in a lot of negotiations if they came into situations in which both cars were approaching each other - particularly if they had cars behind them.

There are a fair number of roads like this in residential massachusetts (Medford and parts of Somerville anyway). Basically one car pulls over in front of a curb cut or where no cars are parked and lets the other car go by.

There's some single-lane Swiss mountain roads, with occasional passing spots. There's a specific rule in the Swiss driving code that the car going up has priority, so the car driving down gets to reverse to the closest passing spot - presumably reversing upwards is less scary, as guardrails may or may not be present...

But then there's the extra rule that on some roads if you meet a "car postal" (bus service run by the Post Office) it gets to decide what you do :)

Interesting. I think that's the same rule in hiking as well.

Same uphill/downhill rule here in California. Besides being less scary to go uphill in reverse, you're a lot less likely to get stuck.

This happens all the time.

Generally we look at each other. We make an immediate judgement on who is the better person at reversing and where the last passing spot was. Then one person will reverse while the other thanks him.her profusely.

So, quick rule:

1. The man reverses.

2. If two men meet, the man with the shittier car reverses

3. The younger 'boy racer' will reverse and show off their superior skills.

This gets super complex when two busses meet, both generally have a line of traffic behind. This is when people have to get out and talk. Ugh, this is the worst. Fuck busses.

I suppose this is some sort of game theory, as both partied want to pass as soon as possible. Neither wants to go out of their way too much. They will both come across this situation many times so are pretty much resigned to it happening and that some of the time they will have to reverse. Anyone who doesn't like this will drive to a better road and take a much longer route.

Single lane one-way road you mean?

There is a few miles of roads like this near me:

https://www.google.co.uk/maps/@50.2513013,-5.2627719,3a,75y,...

Sorry for the huge URL, but as you can see it is a two lane road where one lane is blocked by parked cars. To negotiate this you drive past a few parked cars then pull into a space. You can't necessarily see a space to pull in, but if the car that is trying to pass you stops you can assume it has stopped in a place that will let you past. There is often a lot of reversing and swearing.

You can see this clearly ... in the distance is a blue car, the driver has noticed that the Google car is coming towards it and has decided not to enter the narrow section until the google car has exited it. Some people will do this, others (as the blue car has priority over the google car) will drive on, making the google car pull into on of those gaps on the left. If there is no gap, google car must reverse and find a gap.

I can see us having a lot fewer problems like this in a world with far fewer parked cars.

If you drive around 30mph on average, and hit 100,000 miles after 6 years, your car has spent over 90% of its life parked somewhere.

That doesn't solve the problem for cars sold today though.

> They're getting their money's worth out of that billion dollar acquisition last March. I think Kyle Vogt is GM's MVP.

Many people, including me, were critical of the acquisition; acquiring an year-old company barely having any technological breakthrough for a billion dollars sounded like a folly. Slowly, it has become clear that the impetus behind the deal wasn't necessarily acquiring the technology but bring an ardent person on board who can accelerate GM's self-driving efforts. I have a feeling this deal would cast itself as one of the pivotal moments in the company's history.

Very impressive, I'm wondering how it was able to get around the white truck that was double-parked? Does it have sensors on the mirrors and point straight out? I imagine it would have no information as to oncoming cars, with the truck blocking almost everything, so I wonder what the algorithm is for that.

Tesla posted about updates to their auto pilot system which included bouncing radar under the car in front of them.

https://www.tesla.com/en_AU/blog/upgrading-autopilot-seeing-...

At around 1:54 in the video, it comes to a stop light that is in front of a Shell station's gas price board. It actually took me a second to notice the stop light. There is another light on the left side of the screen, but it still made me think.

Is it likely just using some image recognition and a confidence level to determine when it is safe to go?

If so, could the system be fooled by a sign that is meant to always look like a red light? Or worse, a green light to trick the system to run a red?

Wouldn't the average human be fooled by those things?

Please don't get driver's license.

The system is likely capable of distinguishing a lit green light from a sign painted like one. If you're simply asking "would the system be fooled if someone nefariously hung an always green light where a red one should be" surely that is a circumstance that would fool a human as well.

>The system is likely capable of distinguishing a lit green light from a sign painted like one.

I doubt it. It's not even clear that it shouldn't treat a painted green dot like a green light.

I'm not convinced it would distinguish the two.

Perhaps the most concerning thing is that someone could create a nefarious image that's not obviously nefarious to humans. I mean, we've all seen the very slight adjustments required to trick state-of-the-art image classification systems into thinking a cat is really a crocodile.

Eventually, I expect we'll have some national standard for wireless transmitters on traffic signals, so they can send a message (timestamped and cryptographically signed with the state's department of transportation's public key) to the cars to tell them unambiguously when to stop and when it's safe to go.

(This could help human drivers too; just add a red/yellow/green light to the car dashboard for those times when you can't see the actual light because there's a big truck in the way, or the sun is directly behind the traffic signal.)

I also expect that by the time this sees near-universal adoption, we'll already have the image recognition corner cases mostly worked out.

Las Vegas has some communication from traffic signals to vehicles [1]. I can't tell from the article how it is implemented; it might just be that the city's traffic control computer sends traffic light information out, instead of individual lights communicating.

[1] http://www.theverge.com/2016/12/12/13923254/audi-v2i-las-veg...

A nitpick, but you would sign with the private key, and verify with the public key.

Thanks, that's what I meant.

How about "Signal Ahead" light boards on the back of semi trailers, which relay the traffic signal to the cars behind. Like Flashing Rear End Devices on trains, which operate by telemetry and function as red signals.

Ought to be maintenance free and a nice way for the big trucking lines to differentiate their public face - similar to those aero ground effects that are showing up on trailers.

Possibly, but I imagine the incentive isn't really there. Someone intentionally doing that would be penalised and for minor gain. Someone could do it now to fool human drivers and it hasn't really happened.

e.g., stand in hi-vis at an intersection holding a "Slow" sign regardless of the red light above.

Another good question is what does the car do if the stop sign or traffic light is obscured by trees?

I swear to good the city of San Francisco could car less about this issue. I've gotten to the point where I look for the painted STOP on the road, rather than the stop sign.

I assume either the computer could recognize the painted letters on the road (they are often in bad shape) or the GPS would inform the computer that a stop sign exists? So it doesn't have to look for it?

Seeing a firetruck in this made me wonder, how do self driving vehicles deal with emergency services behind them? Detect the flashing lights?

Maybe the real question is "do they?" - maybe they don't? Or maybe it's a problem being worked on. I can't say I've heard of anything in this domain; it might be one of those problems/challenges which could be worked on, a paper written up, and some kind of solution presented that might net a win for whomever tried.

Given all the crazy corner cases I know they have covered I cannot imagine they don't deal well with emergency vehicles. For example, I know Waymo has put a ton of work into just covering the cases when motorcycles are lane-splitting and overtaking them at high speed [0]. I'm sure they deal with ambulances just fine.

0 - http://extremepowersportssa.com/google-self-driving-car-proj...

Not to be pedantic, but it's lane sharing, not lane splitting. Lane splitting, aka riding the dots, is illegal even in CA

Illegal or not it totally happens all the time, so I assume they are engineering for it.

No, this is legal in CA.

California Highway Patrol guidelines: http://lanesplittingislegal.com/assets/docs/CHP-lane-splitti...

Well, we're probably missing the point, and you probably both know, but I'll continue anyway.

The reason lane sharing is legal in California is because there is not a prohibition against multiple vehicles in the same lane adjacent to each other. The other states, one presumes (since we know it's only legal in CA), specify that vehicles may not drive side-by-side in the same lane.

I'm not sure what CA law has to say about riding between lanes (or, probably more accurately, occupying 2 lanes at the same time which would be what you were doing if your wheels were going right down the line), but all this business is why we call it lane SHARING; in theory, any time spent on the white line is during a lane change (properly signalled, of course, except for those who want to point out that CA state law is pretty vague on whether you even need to signal a lane change).

So the fact that the CHP guideline calls it splitting does nothing to change the legality of splitting versus sharing.

But the motorcycle community typically prefers to call it lane sharing, as that's the circumstance under which it's definitely legal (assuming you are not breaking any other laws at the same time).

And yes, I saw a demo from Google a number of years ago where the car very gracefully handled a motorcycle lane sharing past it.

Your comment matched my understanding from ~10 years ago (from discussions on CA motorcycling forums). However, AB-51 passed last August and it's added "lane splitting" to the vehicle code (which I just found out about 10 minutes ago).

https://leginfo.legislature.ca.gov/faces/billTextClient.xhtm...

Anecdotally, I'd heard of motorcyclists cited for violating the CA Vehicle Code if they occupy two lanes: "Existing law requires ... that a vehicle be driven as nearly as practical entirely within a single lane and not be moved from the lane until the movement can be made with reasonable safety."

This resulted in the semantic distinction. It's not illegal for two vehicles to share a lane. It is illegal (and still is) to drive in two lanes or to drive "between" the lanes.

Literally every motorcyclist I know here in California calls it lanesplitting or just splittin'. We're not very good at sharing which is why we are riding single person vehicles in the first place.

Thanks for the follow up! This looks to be a new law. For years it was always lane splitting illegal while lane sharing OK.

I've been riding for decades and didn't know about this change. I wonder how long it will take the cagers to figure it out. Hopefully the autonomous software does first!

http://leginfo.legislature.ca.gov/faces/codes_displaySection.... begs to differ.

The definition of "lane splitting" there is not what GP referred to, so this is just equivocation.

Totally legal in CA.

Technically riding the dots is still illegal in CA even with the new "lane splitting" law.

If you don't ride, you see everything as lane splitting. If you ride, you damn sure know the difference between lane sharing and riding the dots.

Reminder to never bring up motorcycle topics of discussion to people who don't ride.

There's jargon and terms that cagers just don't get. (Example)

Or just take your audience into account and use language that will be commonly understood or take the time to explain terms that might cause difficulty.

I think GM is in a great position to compete with Uber and tesla if they capitalize on the tech.

As a customer I'll still be happy with multiple choices driving prices down and offering a better service than buses and taxis.

God I hate buses in the US. Public transport is very broken.

I'm sad to say this, but I think GM is going to crush Tesla in the long run. GM knows about building cars, tech becomes cheaper and available over time. It will take Tesla longer to learn how to build and mass produce at a cheap price than for GM to have all the tech they need to match up with Tesla. Initially I was worried that Honda would be the one, but they don't seem to be taking the electric and self driving car serious. GM has their volt/bolt with decent range, I happen to work at a company that shares a lot with some of the auto manufacturing suppliers for American car companies, and I see them testing their self driving cars sometimes. Oh and they have a bunch of Teslas for inspiration I suppose. :)

I think GM is going to crush Tesla in the long run.

Maybe in the short run.

Jan 2017 EV sales:[1]

    Chevy Volt:         1611
    Toyota Prius Prime: 1366
    Chevy Bolt:         1182
    Tesla Model S:       900
    Nissan Leaf:         772
    Tesla Model X:       750

The Bolt has only been shipping since late December. That's just getting started.

[1] http://insideevs.com/monthly-plug-in-sales-scorecard/

Why do you say so? That's 2793 Chevys vs 1650 Teslas.

There are more people who can afford Chevy than Tesla.

Tesla hasn't demonstrated that they can manufacture cheaply, they have promised and we are still waiting. GM might not have a luxury electric package, but they know how from their experience in building Cadillac. Tesla not in the dealerships is a problem. GM manufactures outside of the country and has about 30 factories. They can build and sell electric car outside. Tesla doesn't, building, shipping & custom fees increases price of car a lot.

I'm not convinced that Tesla will win this battle if all variables remain the same. They need to move faster to have a chance.

Animats is saying you don't have to wait to see GM beat Tesla, not that they won't beat them in the long run.

Of course, comparing sales of Chevy's and Toyota's mostly low-end lineup to Tesla's luxury offerings is pretty odd anyway. At the moment it seems pretty TBD.

(I also maintain that including cars like the Volt and Prius in the same category as the Bolt and Model S doesn't make much sense. You may as well compare the Bolt and Model S to all gas-powered cars if you're going to include those gas-powered cars in the mix — it's equally unenlightening.)

Allegedly some large fraction of driving takes place within 20 miles of home. In light of that, including vehicles with 20-40 miles of EV only range makes perfect sense.

Sure, a lot of driving is done near home. But if trips outside that range were really negligible, gas tanks wouldn't be a standard feature in those cars. They're not including them just for funsies.

As an aside, I think you're slightly misunderstanding that statistic. If I do a 60-mile round trip directly away from my house and back, 66% of those miles are within 20 miles of home even though the trip isn't anywhere near 20 miles in total. If I drive 15 miles away from my house and then drive a a semi-circle for 80 miles, all of those are within 20 miles of home.

It's worth noting that Tesla doesn't report monthly sales numbers.

To be fair, the Model 3 is not available yet, so I'm not surprised to see cars like the Volt, Prius and Bolt selling better than 70k+ cars like the Model S and X.

> The Cruise car makes a left turn and then gets stuck behind a delivery truck. It waits as a couple of other vehicles overtake it

I suppose I'd probably notice if I was a semi-engaged driver, but not as much if I considered myself "just a passenger" in an autonomous vehicle. It looks like barely more than 30s elapses there.

I'm not sure I would have done anything different than the computer did. I'm a cautious driver and have been in similar situations where I felt it was prudent to take a moment and really try to see past the truck to look for any oncoming traffic. I've then had an impatient driver behind me decide to zoom around.

This is a good point, I'd want my car to be more like you than me, where I'm the guy impatiently going around people who hesitate for too long.

Just because it took it's time to navigate the problem it doesn't mean the decision was wrong. If you're in a rush you could easily take over the auto-drive feature. But otherwise let it auto drive in cautious mode for 99% of the time.

I used to be the type of driver who would impatiently zoom around, until one day I noticed several deer beginning to cross the freeway. I slammed on my brakes, but the driver behind me honked and swerved around me and, at over 80mph, took two of the deer with him. He left the scene in an ambulance and his car went to the scrapyard on a flatbed. No one ever told me what happened to the man, but while we were waiting for the ambulance and while he was being loaded, I never saw him even wiggle a toe or blink an eye. He was breathing but completely unresponsive. I don't know if he survived or not.

It's moments like that one that can change your outlook on life in an instant. That was four years ago and I haven't intentionally broken the speed limit since then. Even if it means other impatient drivers are zooming around me. I have, however, been rear-ended twice since then.

> Even if it means other impatient drivers are zooming around me. I have, however, been rear-ended twice since then.

Wait, so you already saw someone be seriously injured, and since then, you've caused two other accidents (which are a danger to you too), and you're still going with "I'll drive the speed limit and not drive with the flow of traffic?"

smh

Drive with the flow of traffic: save lives.

That's a bold assumption about his accidents being his fault. Maybe he was rear ended at a light when fully stopped. And even if he was going slower than traffic, it's still the other driver's job not to hit other cars.

It's possible to increase accidents and save lives at the same time. If the average speed of accidents is greatly reduced, that could save lives.

The first time I got rear-ended I was coming to a stop at a red light, the second time I was in the left-turn lane and came to a complete stop at the blinking red light. At no time was there any "flow of traffic" to speak of. The only way I caused those accidents was by not zooming through the light that was changing to red and not driving into the middle of the intersection on a blinking red light.

It's not my fault other drivers don't expect me to follow the law.

> I used to be the type of driver who would impatiently zoom around

I've never been a very aggressive driver, but I've dialed it back even more when I noticed I'd generally meet the people who overtook me in a huff a few miles down the line. Shit's not worth the stress or the fuel burn.

I wish they released a realtime cut of this. I feel like some of the quirks have been obscured with sped up footage.

That video was impressive. It shows great situational awareness.

Here's what I wonder about the future. Is there a point in which consolidation will take place in the auto industry as a result of superior self-driving performance and capabilities? Meaning -- can the value created by self-driving software be more of a factor than the other existing assets of a GM, Ford, Chrysler, Toyota, etc?

I predict that self-driving technology will become a commoditized box + sensors for use by different car brands. It will very quickly reach the point where one autonomous driving engine will not be noticeably better or worse than any of the others on the market.

Interesting. So price will not factor in? And, will a luxury car be expected to have better self-driving awareness/features than a cheaper car?

And before it hits commodity state, what if Waymo's Chrysler minivans and GM's Chevrolets are far superior self-drivers than say, BMWs & Mercedes? How much does that alter the demand curve?

I don't expect there to be different tiers of autonomous driving, because the sensors are cheap and the software can be duplicated at no cost. There are a billion cars in the world, so whatever the initial cost is to develop the self driving software, it's going to be pennies per car in the long run. It will be very tough for car manufacturers to defend their customers dying in a totally avoidable accident because their driving AI -- like a chess engine -- deliberately made a worse move because the customer didn't buy the premium package.

I expect premium car brands to charge more for the same self-driving technology, though, as at least in the short term self-driving capability will be sold as an expensive extra.

I think the commodity stage will be hit two years after the first truly self-driving cars are sold. So it shouldn't give one brand a lasting advantage.

there might be some added value in premium cars - for example certification that it can go 100kmh fully autonomous (better processing hardware, more & expensive sensors), 130 (as per most EU countries) and so on. BMW can have a gap on safe driving at 200 kmh, ferrari 250 kmh.

but overall I feel those premium brands will have very tough times - most of their added value is in driving experience (for example what BMW delivers in ie 3-sedans is a wonderful experience) and this will be removed. not that many people feel the need to pay premium for something that will feel like a commercial airliner, travelling will be race to the bottom costwise.

> most of their added value is in driving experience

I think BMW's added value is in comfort and prestige. People buy a BMW because they can afford it and other people can see that they have one. The ride is smooth, the interior is nice, it has good sound insulation, etc. A small fraction buys them because they are fun to drive. BMW makes luxury sedans, not sports cars. The top of the luxury sedan market is not Ferrari but RR and Bentley where people literally buy the car and pay someone else to drive it.

...will a luxury car be expected to have better self-driving awareness/features than a cheaper car?

Haha it will be the first time in history for BMWs to be driven in safe responsible fashion...

So, silly question - why are these technologies not being made broadly available as driver assist tools while we work our way up to L4 driving automation?

If we're really concerned about making driving safer, this seems like a good compromise until we get to broadly available (and affordable) L4.

> So, silly question - why are these technologies not being made broadly available as driver assist tools while we work our way up to L4 driving automation?

Well, they are. Lane departure warnings, auto braking, self parking are all incremental automation/assistive technologies that have been rolling out slowly and will continue to.

Advertising stuff now is a kind of marketing aimed at hiring and in building enthusiasm for a company like GM that as seen as a has been (not picking on GM, they are simply the topic of this article!).

I think your real question is, "why aren't I seeing these features on current model year cars?" It's about the structure of the car industry. For example, when I worked with <car company X> on brake-by-wire in 1996 it was for the 2004 model year. They had a team with functional steer by wire -- I still haven't seen it in production! This is a combination of historical industry practice, a focus on safety, and regulatory constraints. A mixture of good and bad.

> I think your real question is, "why aren't I seeing these features on current model year cars?"

Or as 3rd party modules which can be installed by a moderately competent mechanic.

And while yes, some of these are being offered, they are far from freely available and affordable - two things which would seem to me to be important for full blown automation.

The 3rd party mod was Cruise's original business model, right? (At least nominally.)

Look up Comma.ai. They got a cease and desist letter from the NHTSA.

... that'd be the guy who refused to answer even basic safety testing questions though, right?

Or am I thinking of someone else?

Freely available and affordable.

https://github.com/commaai/openpilot

> They had a team with functional steer by wire -- I still haven't seen it in production!

Infinity has steer-by-wire in production now.

We have the basic optical and radar lane assists and similar on current models, but as long as I have to keep my hands on the wheel I'm not going to want to have (or pay for) spinning lidars and other expensive tech that I assume this requires.

I think having partially self driving cars that require active driver engagement are a huge safety problem, because if drivers are able to just sit there and not doing anything 90% of the time, they just aren't going to pay attention ever. Some driver aids like automatic braking are probably fine, but 'getting from a to b' is probably something that shouldn't be available until it's as reliable as a human driver.

I'm thinking less "90% automation" and more "Jeeves from Ironman's Suit". Something to help make the driver aware of anomalies from when they're normally under highway hypnosis or futzing around with the radio.

After driving around in Montana today, I'd love for something to be able to tell me with a high degree of confidence if that's really water or ice on the road. Or what speed that car in front of me is going - have they decided it's ice and are driving 30mph slower than I am? Or is that car disappearing into my blind spot going to merge right into me?

The answers were, mostly water (thankfully, see 2 and 3), yes they do think its ice, and yes, he is going to.

When in doubt, slow down. Jeeves would tell you the same.

The problem is that such drastic differences in speed create accidents. Someone going 40 on roads where everyone else is going 75 is going to create emergency braking, panic merges, and a bunch of other potential problems.

Jarvis telling me to slow down when the surrounding traffic is going 75 would not be useful, and would (if followed) make me more prone to be hit or cause other accidents.

Instead, what I want Jarvis to tell me is "this car <highlighted in a HUD> is going 30mph slower than you." This is useful, actionable data that a vision or radar system could provide.

I would think there would be a huge market for something like this as a chaperone for teen drivers.

They need reminders to check blind spots. They haven't built up experience in picking out subtle signs from other drivers.

This is also an issue for other autonomous systems where there's a human in the loop. The solution for that is well known: dial down the automation every once in a while and ask the human to do something, so that they stay aware of their surroundings and don't doze off.

Unfortunately, a car that turns off the cruise control, asks you to recenter the car in the lane or increase the distance with the car ahead wouldn't sell that well, as people would consider it defective.

Isn't this sort of the Tesla approach where they slowly increment improvements in certain settings?

Years back when Google started testing its cars it gave their employees the chance to use their very preliminary technology during their commute. They made it explicitly clear that the driver still had to be ready to take absolute control of the car at any moment but what they found was that their employees were doing the absolutely the opposite: playing on their phones, reaching into the back seat, etc. That's why Waymo has been adamant that they'll only release their tech when it's 100% ready.

> Misconception 1: Driver assistance systems will evolve gradually into fully autonomous cars

> [...]People often argue that such assistance systems need to be supervised by the driver. This makes sense for assistance systems that operate for a few seconds or minutes (such as a parking assistant) but it can not work for systems that drive continuously. Humans are not capable to maintain the state of alert for hours and hours which would be required to immediately counteract possible deficiencies of a driver-assistance system or to take over from it in a split-second.

> We can only entrust the driving task to a driver assistance system when we are sure that this system can handle all situations which arise suddenly and require immediate reaction. This means that driver assistance systems operating continuously on a highway need to be able to cope with rare situations including pedestrians and bicyclists on the highway (they do appear sometimes on highways), accidents unfolding, animals, sudden rainfall etc. Gradual evolution of such systems is impossible; they need to be extremely capable from the first day on which they are put into operation.

http://www.driverless-future.com/?page_id=774

I'm not expecting them to evolve - I'm wanting those technologies that exist today to be used to help drivers today - not when "we are sure that this system can handle all situations which arise".

I assume there's a cost issue? This stuff was expensive to develop, and kind of a luxury, so they probably want to charge as much as possible, which then delays adoption.

At 1:04 [1], it stops for a few seconds before the intersection. Seems to be due to the white car that is parked in opposite direction as all other cars so it waited for ~20 seconds to confirm that the car is indeeded parked? Should a no-motion indicates it's not moving? Why wait for so long?

https://youtu.be/Vfgjemwc9NU?t=63

They marked it with the 'Scheduled Stop' in the top left, and it seems like the passenger is fiddling with some controls as well? I assumed this was to simulate a pick-up/drop-off, or was just a waypoint in the trip?

They included an annotation that says that it was a "scheduled stop".

Duh, didn't see the big "Scheduled Stop". Thanks for the replies.

I was looking at that too until the others mentioned the scheduled stop! I thought it was because the car thought the pedestrian was too close or something and wanted to wait until she crossed the road.

While certainly impressive, keep in mind that it's only one video showing one of possibly thousands of trials with possibly thousands of difficult corner-case situations where the car got stuck or acted improperly. One video doesn't prove anything, especially not one with a person ready to take over at any moment.

Voluntary data of corner-cases where the car got stuck or acted improperly are available from virtually every company in the self-driving car space except Uber. See the "Vehicle Disengagement Reports" at the DMV's website: https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/disen...

American roads look really unfamiliar to someone from the UK. You stop a lot! So many stop signs.

Yeah, we don't really have roundabouts or yield signs at intersections. Though, in SF there's enough cars and pedestrians that I'm not sure either of those would be faster anyway.

That's what you get when the local concerned mothers association sends a weekly flood of crap to the local politicians and newspaper about how "cars go too fast on a particular road... something something, think of the children". How much of that crap is present varies a lot from city to city and town to town. Generally speaking, the poorer the town and the lower the population density the fewer "resident sponsored improvements" to traffic flow you'll see.

There isn't really any planning. San Francisco never removes stop signs, but just adds them steadily as neighborhoods request them. Eventually they are turned into traffic lights. (I imagine "traffic signals installed" must be a performance metric for someone.)

A downside of having stop signs at almost every corner is that drivers seem to forget how to drive in areas without them. They assume that if they approach an intersection without a stop sign then they can drive through it at full speed, instead of looking for cross traffic and pedestrians.

Looking at this space, I honestly expected more drive by wire and remote VR pilot type intermediary solutions than full autonomy. I wonder if remote control will be completely bypassed or if certain sectors will still benefit from this.

How does this compare to other players right now (ex. Waymo, Tesla, etc)?

Car and Driver says Waymo is way out in front.[1] The 2016 California DMV autonomous vehicle disconnect reports are out, and Waymo is doing three orders of magnitude better than the others in miles between disconnects. Waymo is up to 5,000 miles between disconnects, out of 600K miles driven. Waymo reported zero disconnects on interstates and freeways in the last year. For Waymo/Google, this is about 2x better than last year. That was 2x better than the previous year.

Everybody else is reporting orders of magnitude more disconnects per mile.

[1] http://blog.caranddriver.com/in-the-self-driving-race-waymo-...

I'm not sure it's necessarily a good comparison, though. As I recall, Waymo's testing is mostly freeway and suburban driving. I'd be curious to see how much of their testing is done in a heavily urban environment, such as downtown SF or Oakland, and how it affects their disengage numbers.

Waymo's testing is mostly on suburban streets in Mountain View CA, with another test unit in Austin. Their accident reports are usually from Mountain View, and it's almost always them being rear-ended.

Rear-ended is probably a good thing, since that means automation is erring on the side of caution and applying braking when necessary or when uncertainty is above a certain threshold.

Of course someone could make the argument that excessive braking is dangerous, but honestly, the shit I see drivers do behind the wheel of their car weekly (eating, texting, putting on makeup, shaving), I'd prefer excessive braking over any of that.

Many of Google's rear-endings occur at intersections where there's obstructed visibility to the side. The autonomous vehicle moves slightly forward into the intersection, detects cross traffic, and stops. Someone behind them then hits them at slow speed.[1][2]

[1] https://www.dmv.ca.gov/portal/wcm/connect/4a39c1b9-ca1f-4184... [2] https://www.dmv.ca.gov/portal/wcm/connect/69feb1fa-ff90-4b7a...

Sounds like it needed to check for tailgaters before braking, as humans must do in most city driving. This must be improved by now.

Driving like a moron is unsafe times 1. Driving to the letter of the law and as timid as possible is unsafe times how ever many people get fed up with your crap and do something dumb to pass you.

It's roughly analogous to doing the speed limit in the left lane.

As I was watching this video I thought about how much further along with self driving cars we could be if the data was open source. Sharing this type of information between companies would increase everyone's velocity and get us that much closer to real world driverless vehicles.

At some point the vehicles will have to talk to each other anyway (ie. car-to-car network), so we might as well formalize it.

This is still pretty simple stuff it has to deal with. If you really want to test Cruise, try driving in downtown SF on a weekday around 5pm.

Or any other location in the US also! Try a Boston winter, or Houston's unending road construction. It feels like we're creating self driving cars that are experts at driving in SF.

Gotta take the first steps. Its ok to keep raising the bar.

NuTonomy is currently testing their self driving cars in Boston.

Let's do a small road trip around CA starting and ending in SF, avoiding the tolls and seeing some nice scenery.

Head north over the Golden Gate Bridge, immediately losing cell service. Decide at that time (no cell service) to make a stop at the Nike missile launch site, passing through the long single-lane tunnel w/o crashing into an oncoming car. Head back to 101, continue north, driving through 3 trees: Chandelier Tree, Shrine Tree, Klamath Tree. Yes, actually drive through each tree. Go back south to Redwood National Park, and then head east on Bald Hills Road, half of which is a gravel road. Pass through Hoopa. Make an unscheduled stop at the overlook to the original Hoopa campsite. Head to Weaverville, pass by the trailhead to Mount Lassen, then to Reno and on to Lee Vining. Stop at Mono Lake, in the dirt "parking lot", for one of the canoe trip tours. Head back north to Topaz Lake, then head west down US 4. Try not to get run off the 1.5-lane road by a logging truck. Head to San Jose and then north up to SF, neatly avoiding the bridge tolls.

For bonus points, do this when the state mandates snow chains for Lassen and/or US 4.

>For bonus points, do this when the state mandates snow chains for Lassen and/or US 4.

If anything that will be easier for the AI. People are the hard part. Fewer people around = easier.

If the AI is tasked with "go there" in deep snow, sand or whatever, it'll do better than your average person because your average person has no idea how to drive on anything but clear pavement whereas the guys writing the code are gonna get the input of experts as well as a bunch of testing and refinement.

People are nothing special. They operate objects that you must avoid. Sometimes they jump in front, and you make an effort to not squish them, but there is a limit to how crazy the person can be before you aren't liable.

Snow is tough. You have to infer where the road is from hints like mailboxes, trees, the occasional plant, and tracks left by other cars. In that area, sometimes you get poles that are installed to guide the snow plow operators, but there are seldom any curbs. This is a really tough machine vision problem, with things like lane markings and even road width needing to be inferred. Visibility distance can be severely reduced due to falling snow, and headlights will make it glint back at you. It is important to predict slipping before it happens, lest you slide off a cliff.

Of course, we combine the two. You must predict when other drivers are at risk of sliding into you, then avoid being there at that time. For example, a driver descends a hill on a road that intersects your road, and he faces a 2-way stop sign. You are free to proceed right through because you don't have a stop sign, but this could be a fatal move.

From what I know, they do. They are located in SOMA and I believe do tests at most parts of the day.

I'm not sure humans are very good at that either - there are lots of fatalities downtown SF...

Until I see autonomous vehicles testing in blizzard conditions, or even a moderate snowfall, it's relegated to the West coast. This eliminates nearly 2/3 of the US population (who live in areas where snowfall occurs).

We're at least 5-7 years away from them testing in these conditions.

Surprisingly, they are already testing in snow conditions.

http://www.dailymail.co.uk/sciencetech/article-3399315/The-d...

That 'testing' is taking place on the University of Michigan's M City track. I'm looking forward to testing during these conditions, in the real world, with someone behind the wheel acting as backup.

I'll revise my 5-7 year comment then - I think it will this amount of time (if not more) for these autonomous vehicles to move beyond testing (prod) in winter climates.

Driving on the interstate in a snow storm with semi trucks is no joke - there might actually be a threshold in which pulling over is the best option.

Pretty sure Volvo is testing in Gothenburg right now. (also, it's snowing there right now)

As in the first video there is a ~30s delay when stuck behind a truck. Which I think is reasonable and most human drivers take far too many risks. Seems to be than humans will learn to drive better thanks to self driving cars :).

Given GM's partnership with Lyft, makes me wonder if Uber should be concerned.

Looks like the car cuts off another other driver here?

http://www.youtube.com/watch?v=Vfgjemwc9NU&t=0m42s

Looks like both cars have a stop sign, and this one got there first.

My driver's ed teacher would have dinged me if I pulled that. True the car got there a smidge after the POV car, but when turning and going through another car's lane you should always yield to that car before going into the intersection. I think a human driver would have said, "Yup, I got here 0.5 seconds before him... but I'll let him go before I attempt the turn."

At 00:49, shouldn't the car have yielded to the car coming from the opposite side of the street?

Since the GM Cruise car was at the stop sign first, no.

Is this going to be Cadillac's "Supercruise"?

Anyone knows how those technologies work at night time ?

Shouldn't make a difference as, as far as I know, none of the tech they're using is visually based except for a few of the cameras. Most of the primary systems are LIDAR, RADAR, or some other tech that doesn't rely on perfect visual detection and even the ones that do have a mode for "night vision" or a modifier to allow a similar type of function.

Is the tesla "auto-pilot" visually based? From what I recall the fatal accident was because the trailer was a similar color with the sky or something like that so it was not detected as an obstruction.

Parts of it are, yes. The issue with the truck was the height. The logs indicated that the trailer was being ignored as an overhead sign because its height was greater than the height of the cameras. The visual camera data was ignored completely by the system as the truck was perpendicular to the direction of travel.

Even lane detection ? I don't see how signs and paintings could be detected with radars

It's not all detected with radar, just the primary systems are. There are visual systems but they're not the primary detection system. With lane lines, it's mostly visual but they filter the input and have location-based data for things like signs and signals.

What about street signs, road markings, and even signals? That must be visual, and they certainly look different at night

Can you create a "normalized" image (hardware- or software-wise) that is independent of the time of day? Is that what you mean by "night-vision" and "modifier"?

I would still imagine that certain light conditions (i.e. just around sunset) could make things difficult.

Would be interesting if anyone knows how they handle this.

They do have cameras for those kinds of things but most of them are a backup system and they have the actual signals, signs, and speed limits embedded in the GPS data. If you do a google search, you can find several instances of Tesla's speed-limiting a car going down the freeway where the GPS data was incorrect. The car all but ignored what the signs actually said as long as it could identify that a sign was there.

I'm sure there are light conditions that it can't handle, but there are so many fallback systems that it doesn't have to rely simply on visual data.

Ah, ok, so if the can encounters a pre-mapped street signs, the car doesn't bother to actually check the sign.

Hmm... seems a tad dangerous to me. What if the sign is changed, or two signs are almost in identical position?

I suppose it will be ok if the pre-mapped version is used, whenever a confident visual prediction cannot be made (e.g. due to light, mud, bullet holes, or whatever).

This is part of the issue that's being addressed in the current autopilot. Tesla throttles the autopilot speed to be no more than a certain amount past the posted speed limit. Drivers have been complaining that the posted limits don't match the pre-mapped limits already.

.... still seems a bit dangerous.

Would also be interesting to hear what the other teams do, but I only remember seeing videos in daylight so far.

Somehow I completely overlooked this issue in the past.

I think one of the more depressing things is that (potentially) many driver-less cars will be gas powered. Does anyone know if Waymo is going to be used on any electric cars?

There doesn't seem to be any reason driver-less cars have to be electric, but if big companies pushed the tech onto electric cars it'd be a nice way to speed up their adoption and infrastructure for them.

>I think one of the more depressing things is that (potentially) many driver-less cars will be gas powered. Does anyone know if Waymo is going to be used on any electric cars?

It's not like the software isn't portable or that GM has any aversion to building electric if market conditions are favorable.

Of course not, but having the tech tied to an electric push I think makes the whole package more attractive to consumers. Perhaps I'm too tied to a Tesla mindset on this though, of course self driving cars offer a bunch of advantages on their own.

I have a strong sense that self driving will be generally distributed through the US car population before electric only vehicles given the pace of software vs. pace of battery technology. This will drive mass adoption of gas powered self driving cars pushing back the adoption curve for all electrics. One way to stem this may be through tax and insurance incentives, a discount on self driving due to safety, a discount for electric for environmental reasons, and an extra discount for both.

I was under the impression that electric cars had a lower cost per mile, but the electrics have a higher up front cost.

If you have a self driving fleet (eg. Lyft, Uber, GM, etc.), you'd want to optimize for the former, not the latter. Uber has proved that if you can get a car within a couple of minutes at a low enough price, it's a valid enough substitute to car ownership, at least in cities.

The Cruise cars are already built using the Bolt electric vehicle.

Reply on news.ycombinator.com