Even though we've all been thinking about self-driving cars for a long time, seeing videos of that steering wheel turning in the empty front seat is surprising. I upvoted this article just for those few seconds of video.
There's a whole class of things that are surprising when you see them for the first time, even if you think you're used to them. I wonder what properties this class of ideas have. Are they more exciting ideas? Do startups working on such things make more money? Or is the surprise merely an artifact of the degree and/or type of brokenness in our own mental models?
> Even though we've all been thinking about self-driving cars for a long time, seeing videos of that steering wheel turning in the empty front seat is surprising. I upvoted this article just for those few seconds of video.
My biggest surprise was that they didn't remove the steering wheel entirely and let people sit at the front. That steering wheel moving on its own made me very uncomfortable.
Let's say the LIDAR on top breaks down while the car is on a trip and that the car manages to safely pull into a parking lot. You want to get it back to the shop to fix it. You could always send a tow truck, but if it has a steering wheel, it's a whole lot easier to just send a driver.
Let's say the Lidar took down the CAN bus with it. Why needlessly complicate a v0.9 product that came with a mechanically-linked steering wheel from the factory?
I believe they are legally still required to have the steering wheel and have the ability for a human override. What you are describing is the core aspect of Level 5 automation (and a little bit of Level 4). We're seeing the first production Level 3 vehicles this year, so we are still a ways away from no steering wheel (probably ~10 years, depending on when laws are updated to allow for it)
> I believe they are legally still required to have the steering wheel and have the ability for a human override.
That makes no sense when they're not required to have a driver with their ass in the seat, are the elderly passengers from 00:33 going to unbuckle and jump over the front seat to take over driving?
The override may not be guaranteed to be useful in all possible cases, but surely you can think of situations where you'd really want there to be a steering wheel override available?
They actually built cars without steering wheels previously, but then decided they shouldn't be in the car manufacturing businesses. Now they are just repurposing the Pacifica mini-vans.
Yeah I know, I just find it weirds. I guess they'd still have had to go through the DOT if they removed the mechanical driving bits from an existing car and didn't find it worth doing.
Waymo never had any intention of manufacturing their own vehicles. They've been failing at finding a secure manufacturing partner for about 4 or 5 years now. Sebastien Thrun and Sergey Brin went into a number of auto executive board rooms basically suggesting that the auto industry should be ecstatic to become like foxconn, a brandless low-margin manufacturer of hardware to support Google products and services. The response from auto execs was basically "Fuck that noise", and the carmakers are responding aggressively to the shots that have been fired across their bow.
Longer term this is a problem for Waymo. Their supposed partnership with Fiat Chrysler isn't a full-blown, intergrated manufacturing partnership. Waymo is mostly just buying Pacifica mini-vans from FCA and retrofitting them with Waymo's own custom designed hardware at a separate facility in Nova Michigan.
This is not cost effective, and not viable as a long-term strategy. If Waymo does have a strategy for producing robotaxis in the tens of thousands, they've been very quiet about it.
With every year that goes by building an autonomous OS gets easier. Best practices have been established, knowledge continually disseminates throughout the industry, there is, every year less software and hardware that needs to be built from scratch. Automobile mass production, on the other hand, isn't getting any easier. It's as hard as it ever was. So really, over the long run, it's the legacy automakers who have the true competitive moat, and the real bargaining power.
> Sebastien Thrun and Sergey Brin went into a number of auto executive board rooms basically suggesting that the auto industry should be ecstatic to become like foxconn, a brandless low-margin manufacturer of hardware to support Google products and services.
Did… the… what? They didn't actually try that did they? Did nobody tell them auto exec have been fucking over their supply chains for a century?
Yes, and there's a strange interview with the CEO of (Fiat-Chrysler Automobiles about their deal with Google/Waymo. He seems to have a sense of wonder about dealing with Google, because Google is bigger than FCA. He's not used to that.
On the other side, Continental, the huge European auto parts maker, is quietly starting to make all the parts for self-driving cars. They bought Advanced Scientific Concepts, the good flash LIDAR company, and are preparing that technology from volume production. They make the sensors, actuators, ruggedized computers, dashboards, etc.[1] Delco is also active in automatic driving. Self-driving may just be an auto accessory.
They didn't say it in so many words, but that was basically the terms of the deal. In late 2015 former Hyundai America CEO Krafcik was crowned CEO of Waymo by Larry Page, who saw Krafik as a guy who knew how to talk turkey with car guys.
Krafcik's first order of business was to hammer out a deal with Ford, and by that time Waymo had ceded on many of their initial terms, as they were starting to get nervous, because it does take a good 5-7 years to design a vehicle and tool up a production line, and time is of the essence.
Apparently Ford wanted significant capital investments from Alphabet as part of the deal, and Alphabet CFO Ruth Porat said 'no' to that, so the deal fell through and Ford's then CEO Mark Fields was (eventually) fired and replaced by Jim Hackett as a result.
The other thing that happened around the same time was a talent exodus from Waymo. Ford hired ex-Google self driving car project program manager Brian Salesky to lead the formation of Argo.ai, Ford's own robotaxi subsidiary, and Argo seems to be coming along but it's too early to gauge how well they're really doing.
This is why the Lyft/magna announcement yesterday was interesting. Magna kind of already is the auto industry's version of Foxconn. If the car companies aren't leading the pack in terms of autonomous tech, and aren't willing to be the manufacturers for the companies that are leading, they're going to be shut out of this race.
I can understand GM wanting nothing to do with waymo, they've actually got some promising autonomous tech. But FCA should be jumping at the chance to work with waymo, unless they've got some top secret autonomous research happening that's a lot further along than anybody could guess.
GM had a prototype car a long time ago with a 'drive by wire' system (as I'm sure other manufacturers have had). There was a steering wheel but it wasn't connected to any steering column/shaft. It just had sensors that recognized the actions and turned the wheels accordingly.
A system like that could work where the steering wheel doesn't respond to the user when self-driving but is enabled when needed. Prevents accidental steering.
I left my Linux laptop unplugged by accident the other day and it suspended to disk before the battery ran out so I lost no data. It's like Linux became usable and I didn't notice.
Since the change to Vista/Windows 7 required all new windows drivers for all printers, (many of which were old, outdated, and manufacturers didn't want to redo them) Linux has seemed much easier to print with..
I know this might seem like a naive question but, is there any requirement for a driverless car to have windows at all? perhaps a biological or neurological dependency?
Easier to get out of the vehicle in case of emergencies, when the door is stuck, or easier to see which side of the vehicle is unsafe to exit, easier for humans to take over in emergencies when they have vision outside of the vehicle.
Similar reasons apply to any vehicle which is not driven by you. Bus, airplane, train...
It's pleasant to be able to see out, it's important for avoiding motion sickness for many people, and it helps you trust the vehicle is not going to drive you off a cliff or into oncoming traffic. So, no, it's not a hard requirement, but I can't see a windowless car being a success.
What's the engineering tradeoffs required for windows? Would eliminating windows result in a safer, cheaper car?
Would projecting the outside world on the inside of a tin-can car meet the benefits ascribed to windows? (I could imagine some dystopian future where windowed cars are considered to be luxury items).
> Would projecting the outside world on the inside of a tin-can car meet the benefits ascribed to windows
The motion-sickness reducing benefits of staring at the horizon is usually ascribed to the horizon being very far away and not moving along with the car. Until your projections are holographic, it won't help.
It's probably not going to be cheaper in an interesting way.
The up front cost of a vehicle is something like 1/3 of the cost per mile (probably less for commercially operated vehicles), so even a pretty big decrease in the purchase price will only have a moderate impact on overall cost.
No more or less than the back of a bus I imagine. People seem to prefer windows there while being driven, and I don't suspect removing the driver changes that.
I believe back in the 90s there was a plan to have passenger jets without windows as they added weight, cost and reduced fuselage strength. Never got anywhere.
Windows are definitively required by law in many countries, as it allows other drivers (in other cars) to see through a vehicle to anticipate for what's happening in front of it. In my country (the Netherlands) even tinted windows are prohibited for this reason.
Also, for most passengers in a driverless car, windows would be important. As many humans (including myself) get motion sickness if they have no reference of the outside world.
Wonder if it would be cheaper / safer / lighter to remove windows from airplanes and just mount window shaped monitors in their place that display footage from fuselage mounted camera.
It would be cheaper, and also lighter, which means even cheaper. It would not be safer; the reason windowed planes are heavier and more expensive is because they must meet full safety standards while also having windows.
This concept has been brought up numerous times and has never taken off (ha). I don’t think passengers would react well unless it was literally indistinguishable from a window, and we’re a long way from that.
I'd be willing to give up windows in planes if they had the ability to switch to different perspectives around the plane from like a tablet or something. Like being able to have a view from the Pilot's perspective or being able to look directly beneath the plane. That'd be pretty neat. Maybe add the ability to link up some type of account where you can save photos from the flight if you saw something cool.
I certainly wouldn't like it. Flying is largely unpleasant these days for anyone who isn't paying for first class and TSA pre-pass, but one bit of joy is watching the earth recede, and the clouds and landscape pass below. (It's even better in a little Cessna where you can open the windows.) Having the window replaced by a low-quality video screen that will probably soon be showing no-skip ads like the seat-back screens do these days would kill that.
> “The most rewarding things were the yawns,” said Krafcik.
For me, this is the best line in the article. Moving from excitement to normal in a single trip. The "user experience" can only improve from this point.
Yup, and I'd imagine that in their next Uber trip where they are forced to make small talk with their driver about the weather they're going to be missing their robot driver.
I think you run into legal issues when you take the steering wheel out of the car (i.e. it probably is not legal to drive on public roads under the current laws).
I'm happy to see people do that. And thank Alexa or Google Now when they get what they want. Otherwise I'm afraid that a degraded social status for conversational AIs will result in a tiered status system that will eventually include some people.
Not that it doesn't happen now, but I'd prefer to have the trend avoid further tiering.
Exactly. I look forward to being told "Thank you human" when the robot overlords are finished extracting my human electricity to power their batteries.
I hesitate to expect it will have that exact effect, mainly because it will first be implemented in advertising. Unless it has to be said or humans learn the AI will be nicer to them for saying it, I would mostly expect this to be popularized by ad campaigns like this one.
Your fears are inevitabilities as long as AI is used for data mining people’s private lives and advertising, but the public will continue misfiring when they look for who to blame.
I’m both excited that I won’t have to worry about my daughter driving with as many humans on the road, and a bit sad as she already shows quite an affinity for being behind the wheel. Guess dad will have to sign her up for track days.
"Back in the day, people used to do things ourselves. But it was hard, and sometimes people made errors, some that even cost lives. But now everything is done for us. We get driven in our googlecars to our googlejobs where we sit down to watch in case the AI makes a random mistake in doing something, while we watch our carefully curated googlefeeds for news written by algorithm that ensure no nasty "fake news" or unpopular ideas filter into our consciousness.
Then we go home to watch googlesports, in which AI players play the games people used to play, but so much more efficiently than us, while petting our googlecat that lies purring on our lap. (Because, you know pets are a big environmental hazard.)
Waymo's latest iteration can be remotely operated by a safety driver, instead of having one in the driver's seat. It feels more like a symbolic show of confidence in the existing technology rather than an actual technological leap forward.
Remotely controlled isn't the right term. Remotely monitored. The software does the driving, the human steps in if anything goes wrong. From Waymo's prior reports, I don't think the human actually has to step in very often (a few times a year?).
Probably somewhere in the middle. Arizona doesn't get a lot of weather and development is heavily skewed to be new (so the roads are at least more consistent and such).
The dynamic by all these commentators has always seemed to be predicated on the idea that self driving vehicles somehow either need to be universally useful for all driving tasks or they will stay in the lab. In reality the way this stuff is going to roll out is incrementally. Due to the prevalence of on-demand driving services now, self driving vehicles can be deployed literally just-in-time for routes that are known to be safe under current conditions.
It's perfectly reasonable that once an existence proof is shown for one single commonly taken route in America to be handled autonomously (as it already has been) that that is all the foothold needed to start to deliver real economic benefits from the technology, and hence will align incentives for it to, eventually, be perfected.
That sounds an incredibly risky strategy, are you commenting from direct knowledge or just taking a wild guess?
There's literally no such thing as "known to be safe" on a road, a crash could happen right in front of any of these cars, the car promptly kills its occupants through a mistake and then that's the end of self-driving cars for a decade.
Just commenting about the "one crash will kill the industry" part - I've heard this repeated many times over the last 20 years (way before self-driving cars were close to viable). I've repeated it myself.
But it looks like we're wrong. First of all, there have been crashes with even Tesla's autopilot, and it didn't seem to affect anything.
For another, this same logic would've probably meant that driving itself would never have become popular in the first place, or would've been set back many decades because of car crashes. That doesn't seem to have been the case.
I'm not saying these things can't kill an industry, but I think people are pretty tolerant of even unthinkable dangers beforehand, if the incentive is high enough.
Though possibly a bigger backlash could come when a self driving car kills someone in a situation where there are only 2 bad options to choose from. Being that with a human driver nothing was pre-planned people move on. But with a self driving car where such situations have to be decided in advance the backlash could be much more significant.
To expand on your thought... I think "one crash will kill the industry" would have been true if this industry had continued to behave in a cowboy caution-to-the-wind (see early Tesla, Uber) fashion. But it really does seem like cooler heads have been prevailing. There's been (at least the appearance of) openness about various incidents, post-mortems, and plenty of lobbying of both the political and marketing varieties. I think this industry, for better or for worse, has made a nest for itself that will be able to absorb a good amount of setbacks, even of the mortal variety.
People driving have accidents all the time right now that could have been avoided if proper automation was in charge.
At some point if you have an accident and kill someone, and it can be shown in court that if you'd been using the easily available technology that person would still be alive, people will start being prosecuted for gross negligence or worse.
By that logic, why aren't we already prosecuting drivers of older, less safe vehicles? Why don't we mandate that every single car on the road has a breaking system straight from a sports car(while you can still buy cars with drum brakes today).
Surely, if you only bought a more expensive vehicle, you could have saved a life! /s
On the contrary this is an extremely risk-averse strategy since it allows the deployment of autonomous cars to be done in a way deliberately minimizes accident risk on a granular basis. In any deployment scenario of autonomous vehicles there are probabilities of crashes given the inputs to the system: the route, the weather, the behavior of other drivers, etc. When I said "known to be safe" I was using shorthand for "given an extensive array of inputs like route, vehicle condition, time of day, weather, known traffic patterns, prior trip logs, etc, does this route have a sufficiently low Bayesian prior of probability of failure to reach the destination safely to justify the use of an autonomous vehicle, and if so, which combination of system/configuration/etc minimizes the risk?"
Consider the alternative: autonomous cars theoretically get no production use and then suddenly go "on sale" for any buyer to use them in whatever route or condition they want. This would result in an extremely chaotic situation where suddenly there is no opportunity for gradual production deployment to slowly find the edges of their capabilities (which in some cases could result in accidents) -- instead the whole buyer-base of autonomous cars would be exposed to those edge cases, in full, from the moment the cars became available.
edit: Also I'm not sure what you are asking about direct knowledge: I'm stating observations of reality that self driving cars are already being rolled out incrementally in places/routes that they are best suited to.
Yeah, this.
I've argued with lots of people about this, how self driving cars are never going to happen because solving all the problems is too hard.
And yet over the last years we've had improved cruise control, steering assist, emergency stopping, lane following. And now we're getting technology that can't drive in a lot of places but looks useful you a quick routine drive to a local shop on a quiet simple road for example.
It will gradually get more and more useful. One day you'll be thinking why don't we have driverless cars, only driver assist and then realize that you've not actually had to drive yourself for 2 years...
AI is the same. We'll not suddenly invent a general purpose AI like in the movies and books. Instead we'll slowly add features to the tools we have today one by one, until one day we realize that the accumulation of years of small enhancements has made something indistinguishable from AI
To be fair, there was a flurry of non-good news about autonomous cars about 2 or 3 months ago. Actual bad news, in the sense that things had been overhyped prior.
But hey, these things change fast. I have a Dreyfusian[0] prior on autonomous agents (some people think we're on the way to have robotic soccer referees) -- I'm skeptic about what currently passes for reinforcement learning, for one -- but I'm not above changing my mind. Suddenly even.
Actually, Waymo (Google) is behind schedule. They said that we'd already have self-driving cars available to consumers now[1]. They're still running trials right now. There is no commercial service, either through the purchase of the vehicle or taxi-like services.
Nice weather, nice roads. The conditions seemed pretty ideal. Couldn't tell from the video but it didn't look like they were on a road with a speed limit of anything faster than 45 MPH?
Plus, "viable" might not only have been referring to the tech itself. Public opinion and existing regulations, for better or worse, are a part of viability.
I think the answer is probably a little bit of both. Since this is a promotional video, it seems likely that this is in specific areas/times that were less complicated. I mean, take for example the self driving video Tesla made a few years ago [1] remove the legal need for someone in the diver seat, put some passengers in it while driving that specific route, cut it with some upbeat music, and you now have the Waymo video. On the other hand, if it's good enough for specific times and areas, that's still self-driving and still useful, even if not as useful as it could be.
Rodney Brooks, pioneering MIT robot scientist and founder of iRobot, came out this year with a variety of predictions. For what I'd call "real" self-driving cars, robot taxis that can handle what normal taxis do, he says not earlier than 2035: http://rodneybrooks.com/my-dated-predictions/
That's not shocking at all to me. There's a huge difference between "works well in carefully controlled demo conditions" and "works well enough to replace a mature product". A sunny day in Phoenix on a carefully controlled set of streets is very much the easiest difficulty level of driving.
Not all days in Phoenix are sunny, but only 5% of days are rainy, and 0% are snowy. As compared with many midwestern cities, which get precipitation about a third of days. And all of which have snow, sleet, and ice to deal with much of the year.
The current program says: "We’re searching for early riders in parts of the Phoenix metropolitan area, including Chandler, Tempe, Mesa and Gilbert." So it's definitely not the 600 square miles you claim, and I've seen no evidence that riders can go absolutely anywhere even in the test area. And I note none of the video was at night.
Even the article you cite is clear that this is very limited, and that Waymo is intentionally avoiding disclosing a great deal, including where, when, who, and how much.
Yes, this is very cool. But Waymo, as it should have, has clearly picked the easy place to start. For a domain this complex, the 90/90 rule definitely applies.
17 years for 100% functionality in Boston and New York is not that long, really. It's soon enough for everyone from car manufacturers to city governments to be seriously considering their longer-term plans.
Definitely. And I think it's enough time for a lot of people to pass through the stages Douglas Adams described:
“I've come up with a set of rules that describe our reactions to technologies:
1. Anything that is in the world when you’re born is normal and ordinary and is just a natural part of the way the world works.
2. Anything that's invented between when you’re fifteen and thirty-five is new and exciting and revolutionary and you can probably get a career in it.
3. Anything invented after you're thirty-five is against the natural order of things.”
It'll be very interesting to see where it goes. One of the discussions I've had with friends is the extent to which private ownership of cars goes from near-universal to niche. Reactions are all over the map, but my personal bet is that a lot of the kids who grow up with the idea of autonomous cars will never develop that very American attachment to personal car ownership.
We've had self-driving vehicles commercially available for decades now that work in certain locations. When someone says that it is more than 5 years away, they're talking about level 4+ cars that can operate everywhere and are actually for sale, not running a closed circuit trial. Five years still seems optimistic.
We're being shown a marketing video as part of a PR campaign at SXSW that says that everything works perfectly. Yeah, please forgive me if I am a bit skeptical.
This company is closely related to Google (who has, arguably, the brightest and smartest engineers working for it), a company who just recently knocked all of their own bigtables DBs offline for an hour by a "no-op config change".
Show me a third party using one of these cars for a year solid with no software maintenance, unexpected behavior, or human intervention, and I'll happily shut up. I'll even let them do that solely in AZ.
As with many online debates, this one seems to be based on a difference in definitions. Some people take "self-driving cars" to mean "cars without a steering wheel that will drive my drunk butt around in the rain", or even "no more human cars on the road". I still honestly think that this will never happen without radical changes to our roads, cities, laws, and economy. Others view "self-driving cars" as an extension of cruise control that might eventually become fully automated. This can be already be seen in cars by Tesla, etc.
I think the "5+ years before it's viable" crowd is mostly responding to drum-beaters of the former, not the latter.
The data about human drivers already accounts for drunk drivers.
The real question is whether self-driving cars would be less dangerous than human drivers. I don't think the data so far has been convincing, to put it mildly.
As far as I understand, the 5 million miles driven by Waymo were mostly under human supervision, so it does not tell us a lot on how safe those cars really are.
Does anyone here think that a group other than Waymo is in the lead when it comes to short, medium, and long term deliverables on delivering actual SDC as a product?
Whether we think of taxi services or privately owned vehicles, I don't see Tesla, Cruise/GM, Intel/Mobileye, or indeed any other venture which is running actual rides with no one in the driver's seat. Not to mention the disengagement data from California, which Waymo was doing very well on. Granted, overall, all of these teams are being secretive, so it's tough to know where they truly stand. Still, based on what we've seen, I say Waymo is out in front by at least 6-12 months.
When it comes to delivering SDC to the entire globe, obviously there will be a geo-spatial distribution of roll-outs, with sunny/dry/well planned out places (i.e. Arizona, California) getting SDC first, and snowy/wet/poorly organized (i.e. Boston's cowpaths) cities getting them much later. Not to mention, there need to be lax regulations that remove obstacles to local adoption, and super long term we'll see SDC-specific infrastructure improvements to utilize their inherent advantages/disadvantages.
But, for now, it seems to me that Waymo is in the lead and that the SW USA and California will get all of these Waymo tech in Chrysler/other vehicles first.
One thing that gets mentioned is the many thousands of miles that have been driven by Tesla vehicles, giving them sensor data in so many situations that may not have been seen yet by other companies. We may have to wait and see how much of an advantage that actually gives them (if any).
Has there been any report on how much data is actually uploaded back to Tesla? I would be pretty pissed as a owner for uploading all of my rides back for privacy as well as bandwidth reasons.
I have read about the sensor suite on Teslas on the road being different from the suite on the Waymo cars, for example. So, the sensor data from Tesla may be more or less valuable, regardless of the amount of it. Wait and see is certainly the game.
All the people I've talked to who have worked on self-driving cars seem to think that Tesla is working blindfolded by not having LIDAR. Time will prove if they are right, but certainly the biggest jump in self-driving car ability was due to adding LIDAR rather than any software or data trick.
Musk's opinion is that passive image recognition (ie just cameras) is a requirement for full L5 autonomy. If you have to build that, then LIDAR isn't necessary.
You may get a significant boost in ability with LIDAR, but it's possible that over-reliance on it will hamper those companies' ability to develop passive image recognition and they will end up stuck.
That opinion certainly seems possible, and Musk does seem to have a track record of being very intelligent and perceptive, and is known to develop his opinions from first principles. So... I'm generally inclined to believe he's correct.
The problem with Musk's argument is that there isn't any indication so far that the machine learning has tricks up its sleeve that are sufficient resolve all of the ambiguous situations drivers come across. Having direct, accurate measurements of the environment serves as a ground truth that you don't have with just cameras. Humans don't have LIDAR, but we do have a different ground truth: we have a mental model of the world, and can reason about it. It's not just that we know things like 'fast cars take time to stop' or 'he looked at me so is probably aware of me', but rather being able to synthesize based on really complex behaviors we've never seen before, and in many cases, direct communication. All that is to say, there is a lot more to human drivers than their eyes, and for the moment there is nothing showing that we're on the cusp of developing real AI that can address those concerns. I don't think it's just a matter of labeling more data and GPUing bigger models.
So while Musk might be right if we get such AI, it's not clear that it's about to arrive, and even if it does, we can still use LIDAR. In fact chances are that if you have such great AI, additional sensor inputs will turn it from 'Level 5 Autonomous Pilot' to 'Ultra-super human pilot who can preemtively react to situations you didn't even realize were possible in the moment.'.
Of course there are questions at hand:
1. Is Elon right and I'm wrong? Maybe strong(er) AI is right around the corner.
2. Is current LIDAR tech good enough to reliably provide the ground-truth I spoke about?
Elon Musk could easily be wrong. Has been wrong multiple times in the past.
But one of Elon Musk's most underrated traits is his adaptability: if a genuinely better/cheaper technology becomes available or if the current picked solution just isn't working, he'll switch to the better option quickly.
For instance: SpaceX originally scoffed at the idea of vertical landing rockets, thinking that parachutes were just fine for recovery. They switched after parachutes failed and then Elon sent out a company-wide email of Masten Space Systems (at the time, not much more than some guys in a garage) doing this:
https://www.youtube.com/watch?v=H2Yt5L5TlGM
It's not even that hard to change your mind like that. It's just that usually people don't do it because they invested so much emotionally/verbally/etc in telling people they had the "right" solution and don't want to lose face by changing course.
The issue is that Musk committed to the "LIDAR is unnecessary" path when he told users that the 6 figure cars he was selling had the necessary hardware to support L5 automation. There is going to be blowback if he suddenly says "Whoops, you gotta buy another or drop 20k on upgrades if you want SDC capability".
I don't think Musk would spring for LIDAR until it's under $1000. If necessary, they can then find a way to retrofit the hardware onto those who already paid for the full autonomy option (maybe tens of thousands?), and for others, they can wrap the (relatively modest) cost into any future autonomy upgrades. "Enhanced autonomy" is a $5000 option for the Model S, with "Full autonomy" being another $3000. A $1000 LIDAR system definitely fits.
I think Cruise/GM is winning. Waymo has distance, cruise/gm has velocity and acceleration. Retrofitting pacificas is not scalable. You basically have to be or buy a car company to actually deploy a profitable self driving taas at scale (and it still won't be profitable overnight) imo, and that will be true until you can buy self driving cars directly from manufacturers, which will probably happen like ~2 years minimum after the first car manufacturer starts deploying their self driving taas.
Why is this? Because self driving cars require shit tons of wiring (and quality wiring at that) and redundancy, and you want to do that in a plant, not by hand; moreover you can save a lot of capital costs by making your taas car more like a model 3, that is, stripping everything out of it, but you can't get those savings by just retrofitting pacificas. Google isn't playing to win at this point - maybe they don't yet believe it's ready for deployment any time soon. Personally I think there will be (perhaps somewhat small) geofenced publicly available commercial deployments within 2 years though.
Also I would rank tesla dead last, less sensor capability than everyone who's trying it seriously (this impacts data collection ability)... model 3 is a botched attempt a self driving taxi, serious financial issues, there's been a lot of turnover in their self driving team with people leaving for places like argo ai (self driving startup ford bought), and, ideologically, they can only deploy electric self driving cars without hurting their brand (though all self driving cars will be at least hybrids).
I think its more likely that once Waymo has a complete plan they will switch to partnering with existing brands, many of which might not have gotten their own self driving efforts to a point where they can compete.
At least I have not seen any signs of Waymo aiming for car mass production.
> and, ideologically, they can only deploy electric self driving cars without hurting their brand (though all self driving cars will be at least hybrids).
You think this is a con?
As a pedestrian and driver I find it useful to see the faces of other drivers. When walking across the road I try to make eye contract with the driver to make sure they see me. Doesn't seem possible with self driving cars. If a display in the driver's seat could make "eye contact" then display my photo with a checkmark - to say "I see you" I'm be more happy about walking in front.
I think a highly directional light that can be aimed will work better than a display. The car can just point the light at pedestrians that it has recognized.
The hard part is a fast and reliable mechanism for the pointing. I guess an array of lights has the advantage of no moving parts and the capability to show recognition to multiple pedestrians.
If you want to shine a light at a particular spot, just use whatever technologies are used in video projectors. Except for the fact that you are aiming at different distances, it's the same problem.
> I feel self driving cars need to do everything humans do and then 10x better to succeed
I hope that's not the case. Even if self-driving cars are only as good as the median human driver, a universal rollout of that tech would save countless lives. I really hope that when the first self-driving fatality happens the media or public fear doesn't blow it out of proportion.
There's an opportunity here to do better than human drivers too. An autonomous car will always acknowledge you, and the display for that could be put outside the car, in a way that the pedestrian always sees it. You're not always able to see the face of a human driver, depending on the light conditions.
As a bicyclist, I was taught that just because someone's face was turned in your direction doesn't mean you can tell what they are focusing on. My brother got run into by someone who pulled out after making "eye contact". I go around behind every time, like I'm invisible.
Waymo cars are all over Chandler and Tempe in Arizona. I see many on a regular basis with a good portion self-driving.
It is neat to see all the roads you normally drive in the self-driving videos and it is almost taken for granted that they are here [1][2]. Chandler has always been in early on technology due to many reasons but that lots of programmers/engineers work here due to intel historically, projects like Iridium started here, also one of the first places to get cable/broadband internet in the 90s.
With that said, Chandler is very easy to drive, square roads, good weather, clean roads, clearly marked and kept up, minimal freeway needed etc. It is a great place to start but still lots for the cars/algos to learn, however it IS happening and I see it everyday.
Many of the videos are around Chandler downtown that is a bit busy and has lots of movement/pedestrians[3], so that is good that it is able to navigate that, it is close to busy urban environments.
Let's hope they start bringing these cars to upstate New York in the winter, only then can we seriously consider Level5. The salt layers on so thickly that I would be surprised if LIDAR and cameras could work through that. I am not sure of any solutions besides adding little wipers for all the sensors to keep them clean or adding more repellent paint.
Pittsburgh, where Uber's ATC is based, is another profoundly challenging environment— there are a lot of curvy cobblestone streets with >10% grade that haven't been maintained in decades.
And the weather is often snowy or sleety or otherwise unfriendly.
Google does have wipers on at least one LIDAR design. Another sensor cleaner alternates sprays of washer fluid and compressed air. We had that on our DARPA Grand Challenge vehicle in 2005. Ford has a camera cleaner on some 2017 Ford pickups.[1]
Just remember that the extra sensors are a "bonus": if a human being can successfully drive by just using their two eyes in the driving seat, with perhaps 200° range and a couple of mirrors, then an AI can be trained to do the same.
Has anyone been in one of these? I'm curious about what kinds of mistakes they tend to make, if any, because I assume that's where the learning happens.
148 comments
[ 3.6 ms ] story [ 200 ms ] threadThere's a whole class of things that are surprising when you see them for the first time, even if you think you're used to them. I wonder what properties this class of ideas have. Are they more exciting ideas? Do startups working on such things make more money? Or is the surprise merely an artifact of the degree and/or type of brokenness in our own mental models?
My biggest surprise was that they didn't remove the steering wheel entirely and let people sit at the front. That steering wheel moving on its own made me very uncomfortable.
That makes no sense when they're not required to have a driver with their ass in the seat, are the elderly passengers from 00:33 going to unbuckle and jump over the front seat to take over driving?
Longer term this is a problem for Waymo. Their supposed partnership with Fiat Chrysler isn't a full-blown, intergrated manufacturing partnership. Waymo is mostly just buying Pacifica mini-vans from FCA and retrofitting them with Waymo's own custom designed hardware at a separate facility in Nova Michigan.
This is not cost effective, and not viable as a long-term strategy. If Waymo does have a strategy for producing robotaxis in the tens of thousands, they've been very quiet about it.
With every year that goes by building an autonomous OS gets easier. Best practices have been established, knowledge continually disseminates throughout the industry, there is, every year less software and hardware that needs to be built from scratch. Automobile mass production, on the other hand, isn't getting any easier. It's as hard as it ever was. So really, over the long run, it's the legacy automakers who have the true competitive moat, and the real bargaining power.
Did… the… what? They didn't actually try that did they? Did nobody tell them auto exec have been fucking over their supply chains for a century?
On the other side, Continental, the huge European auto parts maker, is quietly starting to make all the parts for self-driving cars. They bought Advanced Scientific Concepts, the good flash LIDAR company, and are preparing that technology from volume production. They make the sensors, actuators, ruggedized computers, dashboards, etc.[1] Delco is also active in automatic driving. Self-driving may just be an auto accessory.
[1] https://www.continental-corporation.com/en/products-and-inno...
Krafcik's first order of business was to hammer out a deal with Ford, and by that time Waymo had ceded on many of their initial terms, as they were starting to get nervous, because it does take a good 5-7 years to design a vehicle and tool up a production line, and time is of the essence.
Apparently Ford wanted significant capital investments from Alphabet as part of the deal, and Alphabet CFO Ruth Porat said 'no' to that, so the deal fell through and Ford's then CEO Mark Fields was (eventually) fired and replaced by Jim Hackett as a result.
The other thing that happened around the same time was a talent exodus from Waymo. Ford hired ex-Google self driving car project program manager Brian Salesky to lead the formation of Argo.ai, Ford's own robotaxi subsidiary, and Argo seems to be coming along but it's too early to gauge how well they're really doing.
I can understand GM wanting nothing to do with waymo, they've actually got some promising autonomous tech. But FCA should be jumping at the chance to work with waymo, unless they've got some top secret autonomous research happening that's a lot further along than anybody could guess.
A system like that could work where the steering wheel doesn't respond to the user when self-driving but is enabled when needed. Prevents accidental steering.
I listen to way too much youtube when I am working, and am still worried about making the switch.
This may mean more about the progression of cheap postscript-compat printers than Linux.
But the wireless mdns discovery thing works, which is really nice.
Easier to get out of the vehicle in case of emergencies, when the door is stuck, or easier to see which side of the vehicle is unsafe to exit, easier for humans to take over in emergencies when they have vision outside of the vehicle.
Similar reasons apply to any vehicle which is not driven by you. Bus, airplane, train...
Yes, it is a naive question.
Would projecting the outside world on the inside of a tin-can car meet the benefits ascribed to windows? (I could imagine some dystopian future where windowed cars are considered to be luxury items).
The motion-sickness reducing benefits of staring at the horizon is usually ascribed to the horizon being very far away and not moving along with the car. Until your projections are holographic, it won't help.
The up front cost of a vehicle is something like 1/3 of the cost per mile (probably less for commercially operated vehicles), so even a pretty big decrease in the purchase price will only have a moderate impact on overall cost.
Also, for most passengers in a driverless car, windows would be important. As many humans (including myself) get motion sickness if they have no reference of the outside world.
Hence passenger windows in airplanes, even though this is structurally problematic and requires expensive reinforcement around the windows.
Not to mention windows in metro systems that are entirely underground.
It’s just a comfort thing. Removing it would likely decrease customer demand.
Wonder how people would react to that.
This concept has been brought up numerous times and has never taken off (ha). I don’t think passengers would react well unless it was literally indistinguishable from a window, and we’re a long way from that.
* mitigation of motion sickness
* mitigation of claustrophobia
* free lighting
* they're using a standard car body, or they could have not put the mechanical steering wheel in
1. Good weather, sub-freeway speeds. Large parts of the US (and the world) have good weather for most of the year.
2. Going uphill in hot weather, where running the AC would put too much strain on the engine
For me, this is the best line in the article. Moving from excitement to normal in a single trip. The "user experience" can only improve from this point.
Something like Poe in Altered Carbon will be awesome.
Disc: Googler but nowhere close to this project.
http://bigthink.com/ideafeed/googles-self-driving-car-no-ste... Google's Self-Driving Car: No Steering Wheel, No Brakes | Big Think
This is not the last time we will be hearing that.
Not that it doesn't happen now, but I'd prefer to have the trend avoid further tiering.
Your fears are inevitabilities as long as AI is used for data mining people’s private lives and advertising, but the public will continue misfiring when they look for who to blame.
Back in the day, we used to drive without seatbelts.
Back in the day, driving drunk was socially acceptable.
Back in the day, we rode horses.
Etc. It'll be something.
Just like "back in the day, computing was done by humans", "typesetting was done by humans", etc.
"we just drove 'em ourselves!"
"wow, no one died that way?"
"oh no, millions of people died"
Then we go home to watch googlesports, in which AI players play the games people used to play, but so much more efficiently than us, while petting our googlecat that lies purring on our lap. (Because, you know pets are a big environmental hazard.)
..yeah you have fun telling them that.
Are these only working on very specific types of streets and traffic volumes or was everyone wrong and the tech accelerated very fast?
It's perfectly reasonable that once an existence proof is shown for one single commonly taken route in America to be handled autonomously (as it already has been) that that is all the foothold needed to start to deliver real economic benefits from the technology, and hence will align incentives for it to, eventually, be perfected.
There's literally no such thing as "known to be safe" on a road, a crash could happen right in front of any of these cars, the car promptly kills its occupants through a mistake and then that's the end of self-driving cars for a decade.
But it looks like we're wrong. First of all, there have been crashes with even Tesla's autopilot, and it didn't seem to affect anything.
For another, this same logic would've probably meant that driving itself would never have become popular in the first place, or would've been set back many decades because of car crashes. That doesn't seem to have been the case.
I'm not saying these things can't kill an industry, but I think people are pretty tolerant of even unthinkable dangers beforehand, if the incentive is high enough.
Though possibly a bigger backlash could come when a self driving car kills someone in a situation where there are only 2 bad options to choose from. Being that with a human driver nothing was pre-planned people move on. But with a self driving car where such situations have to be decided in advance the backlash could be much more significant.
People driving have accidents all the time right now that could have been avoided if proper automation was in charge.
At some point if you have an accident and kill someone, and it can be shown in court that if you'd been using the easily available technology that person would still be alive, people will start being prosecuted for gross negligence or worse.
Surely, if you only bought a more expensive vehicle, you could have saved a life! /s
Consider the alternative: autonomous cars theoretically get no production use and then suddenly go "on sale" for any buyer to use them in whatever route or condition they want. This would result in an extremely chaotic situation where suddenly there is no opportunity for gradual production deployment to slowly find the edges of their capabilities (which in some cases could result in accidents) -- instead the whole buyer-base of autonomous cars would be exposed to those edge cases, in full, from the moment the cars became available.
edit: Also I'm not sure what you are asking about direct knowledge: I'm stating observations of reality that self driving cars are already being rolled out incrementally in places/routes that they are best suited to.
And yet over the last years we've had improved cruise control, steering assist, emergency stopping, lane following. And now we're getting technology that can't drive in a lot of places but looks useful you a quick routine drive to a local shop on a quiet simple road for example.
It will gradually get more and more useful. One day you'll be thinking why don't we have driverless cars, only driver assist and then realize that you've not actually had to drive yourself for 2 years...
AI is the same. We'll not suddenly invent a general purpose AI like in the movies and books. Instead we'll slowly add features to the tools we have today one by one, until one day we realize that the accumulation of years of small enhancements has made something indistinguishable from AI
But hey, these things change fast. I have a Dreyfusian[0] prior on autonomous agents (some people think we're on the way to have robotic soccer referees) -- I'm skeptic about what currently passes for reinforcement learning, for one -- but I'm not above changing my mind. Suddenly even.
[0] https://en.wikipedia.org/wiki/Hubert_Dreyfus%27s_views_on_ar...
[1] https://www.extremetech.com/extreme/147940-google-self-drivi...
Plus, "viable" might not only have been referring to the tech itself. Public opinion and existing regulations, for better or worse, are a part of viability.
[1] https://www.tesla.com/videos/autopilot-self-driving-hardware...
That's not shocking at all to me. There's a huge difference between "works well in carefully controlled demo conditions" and "works well enough to replace a mature product". A sunny day in Phoenix on a carefully controlled set of streets is very much the easiest difficulty level of driving.
Not all days in Phoenix are sunny and 600 square miles is not "carefully controlled set of streets".
According to https://www.wired.com/story/waymo-google-arizona-phoenix-dri...:
"Once Waymo is up and running in Arizona, it will gradually expand its service area to cover a 600-square mile swath around Phoenix"
By all signs that's going to happen this year.
Not all days in Phoenix are sunny, but only 5% of days are rainy, and 0% are snowy. As compared with many midwestern cities, which get precipitation about a third of days. And all of which have snow, sleet, and ice to deal with much of the year.
The current program says: "We’re searching for early riders in parts of the Phoenix metropolitan area, including Chandler, Tempe, Mesa and Gilbert." So it's definitely not the 600 square miles you claim, and I've seen no evidence that riders can go absolutely anywhere even in the test area. And I note none of the video was at night.
Even the article you cite is clear that this is very limited, and that Waymo is intentionally avoiding disclosing a great deal, including where, when, who, and how much.
Yes, this is very cool. But Waymo, as it should have, has clearly picked the easy place to start. For a domain this complex, the 90/90 rule definitely applies.
“I've come up with a set of rules that describe our reactions to technologies: 1. Anything that is in the world when you’re born is normal and ordinary and is just a natural part of the way the world works. 2. Anything that's invented between when you’re fifteen and thirty-five is new and exciting and revolutionary and you can probably get a career in it. 3. Anything invented after you're thirty-five is against the natural order of things.”
It'll be very interesting to see where it goes. One of the discussions I've had with friends is the extent to which private ownership of cars goes from near-universal to niche. Reactions are all over the map, but my personal bet is that a lot of the kids who grow up with the idea of autonomous cars will never develop that very American attachment to personal car ownership.
This company is closely related to Google (who has, arguably, the brightest and smartest engineers working for it), a company who just recently knocked all of their own bigtables DBs offline for an hour by a "no-op config change".
Show me a third party using one of these cars for a year solid with no software maintenance, unexpected behavior, or human intervention, and I'll happily shut up. I'll even let them do that solely in AZ.
I think the "5+ years before it's viable" crowd is mostly responding to drum-beaters of the former, not the latter.
5 million miles and no fatal accidents. Just 5% of the way to when a _human_ driver (in the us) would expect a fatal accident. Fingers crossed!
https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...
The real question is whether self-driving cars would be less dangerous than human drivers. I don't think the data so far has been convincing, to put it mildly.
As far as I understand, the 5 million miles driven by Waymo were mostly under human supervision, so it does not tell us a lot on how safe those cars really are.
Whether we think of taxi services or privately owned vehicles, I don't see Tesla, Cruise/GM, Intel/Mobileye, or indeed any other venture which is running actual rides with no one in the driver's seat. Not to mention the disengagement data from California, which Waymo was doing very well on. Granted, overall, all of these teams are being secretive, so it's tough to know where they truly stand. Still, based on what we've seen, I say Waymo is out in front by at least 6-12 months.
When it comes to delivering SDC to the entire globe, obviously there will be a geo-spatial distribution of roll-outs, with sunny/dry/well planned out places (i.e. Arizona, California) getting SDC first, and snowy/wet/poorly organized (i.e. Boston's cowpaths) cities getting them much later. Not to mention, there need to be lax regulations that remove obstacles to local adoption, and super long term we'll see SDC-specific infrastructure improvements to utilize their inherent advantages/disadvantages.
But, for now, it seems to me that Waymo is in the lead and that the SW USA and California will get all of these Waymo tech in Chrysler/other vehicles first.
You may get a significant boost in ability with LIDAR, but it's possible that over-reliance on it will hamper those companies' ability to develop passive image recognition and they will end up stuck.
That opinion certainly seems possible, and Musk does seem to have a track record of being very intelligent and perceptive, and is known to develop his opinions from first principles. So... I'm generally inclined to believe he's correct.
So while Musk might be right if we get such AI, it's not clear that it's about to arrive, and even if it does, we can still use LIDAR. In fact chances are that if you have such great AI, additional sensor inputs will turn it from 'Level 5 Autonomous Pilot' to 'Ultra-super human pilot who can preemtively react to situations you didn't even realize were possible in the moment.'.
Of course there are questions at hand:
1. Is Elon right and I'm wrong? Maybe strong(er) AI is right around the corner. 2. Is current LIDAR tech good enough to reliably provide the ground-truth I spoke about?
But one of Elon Musk's most underrated traits is his adaptability: if a genuinely better/cheaper technology becomes available or if the current picked solution just isn't working, he'll switch to the better option quickly.
For instance: SpaceX originally scoffed at the idea of vertical landing rockets, thinking that parachutes were just fine for recovery. They switched after parachutes failed and then Elon sent out a company-wide email of Masten Space Systems (at the time, not much more than some guys in a garage) doing this: https://www.youtube.com/watch?v=H2Yt5L5TlGM
It's not even that hard to change your mind like that. It's just that usually people don't do it because they invested so much emotionally/verbally/etc in telling people they had the "right" solution and don't want to lose face by changing course.
Why is this? Because self driving cars require shit tons of wiring (and quality wiring at that) and redundancy, and you want to do that in a plant, not by hand; moreover you can save a lot of capital costs by making your taas car more like a model 3, that is, stripping everything out of it, but you can't get those savings by just retrofitting pacificas. Google isn't playing to win at this point - maybe they don't yet believe it's ready for deployment any time soon. Personally I think there will be (perhaps somewhat small) geofenced publicly available commercial deployments within 2 years though.
Also I would rank tesla dead last, less sensor capability than everyone who's trying it seriously (this impacts data collection ability)... model 3 is a botched attempt a self driving taxi, serious financial issues, there's been a lot of turnover in their self driving team with people leaving for places like argo ai (self driving startup ford bought), and, ideologically, they can only deploy electric self driving cars without hurting their brand (though all self driving cars will be at least hybrids).
At least I have not seen any signs of Waymo aiming for car mass production.
https://www.youtube.com/watch?v=cdgQpa1pUUE
I guess that's not progress though.
I think a highly directional light that can be aimed will work better than a display. The car can just point the light at pedestrians that it has recognized.
The hard part is a fast and reliable mechanism for the pointing. I guess an array of lights has the advantage of no moving parts and the capability to show recognition to multiple pedestrians.
Your behavior is a hedge against crappy humans that tend to be distracted and not see everything.
If self-driving cars can't avoid hitting people, they are failing the most basic scenario and can't be allowed on the road.
Here's an article saying eye contact is the reason the amount of window tinting is limited. http://www.cbc.ca/news/canada/windsor/ontario-limits-amount-...
I hope that's not the case. Even if self-driving cars are only as good as the median human driver, a universal rollout of that tech would save countless lives. I really hope that when the first self-driving fatality happens the media or public fear doesn't blow it out of proportion.
It is neat to see all the roads you normally drive in the self-driving videos and it is almost taken for granted that they are here [1][2]. Chandler has always been in early on technology due to many reasons but that lots of programmers/engineers work here due to intel historically, projects like Iridium started here, also one of the first places to get cable/broadband internet in the 90s.
With that said, Chandler is very easy to drive, square roads, good weather, clean roads, clearly marked and kept up, minimal freeway needed etc. It is a great place to start but still lots for the cars/algos to learn, however it IS happening and I see it everyday.
Many of the videos are around Chandler downtown that is a bit busy and has lots of movement/pedestrians[3], so that is good that it is able to navigate that, it is close to busy urban environments.
[1] https://www.youtube.com/watch?v=aaOB-ErYq6Y
[2] https://www.youtube.com/watch?v=B8R148hFxPw (360 degree experience)
[3] https://youtu.be/B8R148hFxPw?t=1m50s
And the weather is often snowy or sleety or otherwise unfriendly.
[1] https://youtu.be/e1kLSxv5ovk?t=37