Uber Wasted $2.5B on Self-Driving Cars
theinformation.com
theinformation.com
The most shocking detail:
> Uber’s vehicles had also been having what the company refers to as a bad experience—such as a sudden jerk or a potentially dangerous movement—every one-third of a mile on average.
That's really bad. While there are no uniform standards for what a "bad experience" entails, my general sense is that Waymo and Cruise (widely considered first and second) are doing 10-100x better by that metric.
Another damning bit:
> Uber couldn’t even get the prototype to drive a one-mile stretch between the unit’s two Pittsburgh offices, with the goal of shuttling employees back and forth
I don't even know what to say to that. If you can't make your cars work reliably on a known, one-mile stretch of road, what have you been doing for so many years?
[Disclaimer: I work for Lyft, but not on the self-driving car program.]
Thought Tesla was considered first, in part because Tesla's fleet collects an orders of magnitude of data greater than others.
I’m open to being convinced that I’m not correct here, but it’s really difficult for me to see how Tesla is anything but 1000x ahead of everybody else.
It reminds me of people who build great models on Jupyter Notebook, but have no clue how about the data engineering that goes into serving those models (90% of the work).
Do you own a Tesla? We previously had a car with ACC/ALC and I don’t find it to be even close to the same thing, but that’s me.
Comma.Ai’s device is great. It’s better than anything anybody else (Waymo/Cruise included) has put out there.
Soon as it was in 2016 [1], 2018 [2] in Feb 2019 [3], Mar 2020 [4], or July 2020 [5]?
They have been hyping this up for years, why should I still believe those claims that it will soon be here™? Not only that but they have already advertised and sold "full self-driving" for people that can't use it yet, no beta, no alpha, just doesn't exist and is a paid feature. How far is this from fraud?
[1] https://techcrunch.com/2016/11/19/watch-a-fully-self-driving...
[2] https://cleantechnica.com/2018/11/03/hiding-between-two-anno...
[3] https://electrek.co/2019/02/28/automatic-driving-city-stop-s...
[4] https://electrek.co/2020/03/04/tesla-self-driving-features-m...
https://www.cnet.com/roadshow/news/tesla-model-s-model-x-aut...
Yes you can. At least my ap1.5 s70d does a fair job on quite a lot of roads and I expect that ap2 is better.
> And AFAIK Tesla still can't recognize stopped vehicles on the highway and has no public plan to fix that.
Can you elaborate on what you mean by this? I do have a Tesla (without full self-driving) and it comes to a complete stop when cars in front of me stop on the highway.
One example of a stopped vehicle: https://www.latimes.com/business/story/2019-09-03/tesla-was-...
https://www.wired.com/story/tesla-autopilot-why-crash-radar/
> “ Just before the crash, a large vehicle — a sport utility vehicle or pickup truck — changed lanes in front of him, the driver told the NTSB.”
And how does Waymo handle these scenarios?
The radar returns on stationary objects are discarded AFAIK because of poor radar resolution leading to inability to precisely locate objects both horizontally and vertically, leading to roadside and overhead signs being mistaken for cars.
Common auto radars have limited vertical discrimination, and this is partly intrinsic due to radar wavelengths.
To throw you a bone, I do think that the names of the software lead people to believe that it's something more than what it is. Full self-driving is going to do things like swerve to avoid an obstacle in the middle of the road and it probably doesn't do that right now. But if we go back to the original post, the discussion was about comparisons to Waymo and Cruise. I'm not sure I trust their "more advanced systems" to do that either, and on top of that they haven't made a single production car that works in Erie, Pennsylvania and drives in the snow, changes lanes automatically, and navigates on and off ramps.
Tesla is making progress toward general self-driving, which is far and away more valuable in my opinion. Even if that progress is incremental. I look at it as more of a safety feature and something to make driving on the highway much easier.
If we want to take this a step further, I'm actually not a fan of self-driving cars because it makes it too easy to drive everywhere, and I view driving as a problem. We should be walking.
Hoping the AP rewrite closes some of the gap... we will see!
You don't really need 1000x real examples of a scenario, if one still image provides enough information for you to mimic it in simulation. (E.g. toppled barrel, weathered paint color and reflective stripe, in front of exit guard rail)
Whether you think they're sandbagging or behind the curve is largely a question of whether you're a fan of Tesla.
On a freeway, that's terrifying.
It _seems_ like lots of parallel systems, each with the capability to reduce speed or stop could cope with city traffic about as well as a tourist. but, uh, I guess $2.5B says otherwise.
I can appreciate why the lay person may believe them to be similar, but it's hard to express just how much more challenging even the most basic form of driver assistance technology is in comparison to an autonomous vacuum.
The biggest challenge is speed. Just look at the napkin math of detection range of sensors versus stopping distance of a car. The higher the speed, the more range you need... and lidar, radar, etc. all have a fixed maximum detection range.
The rule of thumb is 50m for radar/lidar depending on how much you are willing to pay for the sensor. To stop the car safely, will take 56m at 100km/h.
Then ofcourse, the rest of the struggle is all in exception handling (getting from 70% to 99%).
Comparatively, the iRobot can use the dumbest and cheapest sensors and still be fully aware of everything around it in its relatively slow path of travel. It also does not need to predict the paths of travel of everything around it. It's enough for it to slow down or stop, especially at such low speed. If it hits something or fails to recognise an obstacle, it's not really a big deal. In comparison, lives can be lost in the car and there will be enquiries and legal battles.
I would put good money on the camera based automation delivering the best product at the end of the day, but requiring an order of magnitude more research and development than a comparitively "dumb" lidar setup. The lidar setup will have better immediate results, but probably will never reach the 99.999% effectiveness that everyone is chasing, and will probably never be safe at higher speeds.
When you drive on the freeway, everyone's going the same way (or at least you'd hope), the configuration of the road is well known, lines are consistent, etc. It's all the little things that need to be accounted for - opposing traffic, pedestrians, parked cars, random obstacles, and any number of other road, traffic and obstacle configurations that exist, that make consistently and safely driving elsewhere difficult for autonomous systems.
I do think Musk was (in part) wrong about LIDAR - I imagine that it could provide a valuable source of supplemental data to be used in training and classification.
So they started Adding curves in the road. Not much, but it improves the engagement factor. Gives the person something to do.
It sounds horrible to have to be fully alert, yet nothing to do.
In programming, I can’t stand shadowing people for more then few minutes before I want to start driving the code myself.
Google/Waymo did a pilot - letting employees commute - and canceled because the system was too good and staff stopped paying attention.
This is a clip, there was a longer clip and description in a Waymo CEO speech in the past year or two.
Most California weather driving is very mundane and boring. ...but the moment you throw an outside variable into the mix - the computer cannot identify objects, and therefore cannot predict outcomes.
We're still teaching them to recognize Stop signs. How are they supposed to distinguish between a floating plastic bag and a 5 year old child?
edit: I say this as someone who very much wants to see this class of science & technology succeed, but I often wonder if we're 2 or 3 Fields Medals away from really getting fundamentally closer to success.
Even the best tech today can only drive reliably and safely in good daylight weather conditions, with no chance of protests.
Pretty sure my wife would claim that I'm worse than that...
Even on identical systems, different parameters will create totally different outcomes.
Maybe it has 'sudden movements' because it's actually a lot after and designed to react to things that Waymo would just plow right through. Maybe they're using crappy non-production sensors. Maybe they are entirely relying on instruments and not indirect data like GPS. I have no idea really, other than to say, these are very sophisticated systems and given the obvious opacity of the work ... it's going to be hard to tell what's what.
This is a scary prospect.
For for that matter, ever 3 or 30 miles would also be really bad in my book.
There are possibly opportunities on freeways but anyone who expects to be able to get robo-Ubers door-to-door for decades is likely to be disappointed.
You're VASTLY underestimating how much progress is being made by companies that aren't Uber.
3-10 years is a more realistic timetable than decades.
"People overestimate what they can accomplish in one year, and underestimate what they can accomplish in 10".
After all, writing an AI is just a summer project for couple of interns.
Here's a video of Waymo CTO Dmitri Dolgov claiming that he's perfectly comfortable putting his children in a Waymo car without a driver: https://youtu.be/fZHDvKw0QTA?t=952. Whether he's bluffing or not, we don't know.
https://battlepenguin.com/tech/self-driving-cars-will-not-so...
Self driving cars are dumping an insane amount of time and effort into computer vision and AI for the wrong problem. For the amount of tax subsidies, our money, given to these companies, we could have put down insane amounts of rail. Hell, we could even do small personal autonomous rail car test programs in some cities and it would still be more useful than the current state of self-diving tech.
Self driving solves the wrong problem, and it's a good 15+ years until we see this vehicles, if we even see them at all.
Human rights (and property rights) can be expensive.
The important metric isn't the cost of the project, it's the cost per km (or mile, if you prefer). The construction costs for NYC in particularly are unjustifiably obscene--it's really on the level of "fire everyone involved and make the new hires go on a world tour of transit system agencies to learn how they do cost control before they do anything."
I believe that's not good enough for Singapore or U.S.
The disfunctionality of the US at large-scale projects emerged in the 80s, possibly to do with both the creation of the modern Republican agenda by Ailes (divisive populism over bipartisanship, and preference of tax cuts over maintaining infrastructure), and the acceptance of the environmental movement by the Democrats. Maybe also by the illusion that the the demise of the USSR meant the "end of history" and the focus can turn from trying to get the US to be #1 to prying to get a bigger slice of the pie domestically...
See "This Is Why Your Holiday Travel Is Awful: The long, sordid history of New York’s Penn Station shows how progressives have made it too hard for the government to do big things—and why, believe it or not, Robert Caro is to blame" https://www.reddit.com/r/gwern/comments/ez0l93/this_is_why_y...
Of course the elephant in the room is money. The public doesn't seem to have enough money to do projects.
Why? Because structurally things cost a lot more than they used to do. ( https://slatestarcodex.com/2019/06/10/book-review-the-prices... -- http://rationallyspeakingpodcast.org/show/rs-236-alex-tabarr... )
One big takeaway point - IMHO - is that there's no scale, no growth in construction. So there's no big market, no new frontier for optimizing it. Sure, there are always a few startups, but that's peanuts.
Where are the trillions of dollars allocated to get good at building things in this century/millenium? Umm, nowhere. (Or basically spent on iPhones, housing and healthcare.) So it's no surprise that everything is custom made, challenged in court, takes forever and costs all the funds available and some.
Conversely, inventing self-driving cars scales really well. Invented once. Commoditized. Sold worldwide for a profit. Massive benefits to everyone worldwide.
It's disappointing that Uber's research isn't paying off, but it's clearly a bet in pursuit of much higher stakes than making a highway.
Even substantially expanded rail will never solve the problem of "I want to get from point A to point B, as quickly as possible, whenever I'm ready to leave" in a fundamentally different way. The transit to the starting station might be shorter, and the transit from the ending station will be shorter, but it's still the same experience.
Autonomous cars solve that in the same way an personal automobile currently does, albeit removing the inconvenience of actually driving.
The former is like-in-kind. The latter is fundamentally-different. Consequently, if I am a consumer, how much would I be willing to pay for the former vs the latter?
Cell phones became ubiquitous because they were a fundamentally-different communication experience ("Call a person instead of a location"). Transit will be the same way, if it changes.
The problem with autonomous cars as compared to rail is that autonomy doesn't solve the other fundamental problem of cars: congestion. If I waved my magic wand and replaced every car in the bay area with a self-driving one overnight, roads would still clog at rush hour.
But whatever it is, you need something major to counter the increase in commuters or else the road system will get swamped.
Additionally, an inherently safe trafic system doesn't require 2 tons of steel around you for protection. The car can be stripped down massively reducing road footprint and ecological impact. You will also hail the smallest (cheapest) vehicle that serves your purporse, a single commuter can comfortably travel in a bumpercar.
We have autonomous trains on controlled lines, and yet they can't run in close formation, and they still need parking space.
There is nothing ordained about this. If we stopped giving away billions of dollars of the most valuable real estate in the world (in the form of parking mandates and street parking) or stopped ruining neighborhoods with freeways and car-centric development then rail (and transit more generally)_could very well be better by these metrics.
But a substantially improved autonomous car driving experience - with cars communicating with each other (actively or passively) to avoid congestion, avoiding parking needs by being on the move always and increasing "bandwidth" by driving close to there other cars and still being safe - may look like railroads on all our roads. Bringing both these worlds closer to each other.
Autonomous cars don't solve the problem of inefficient use of land for transportation (actually, they probably make it worse since there will be a decent fraction of cars moving absolutely no one). Space is at an absolute premium in cities, and people really underestimate just how much space cars take up, particularly since that space is incredibly subsidized.
A train is obviously more still more efficient between fixed points, but self-driving cars are far more economical for less-frequented routes and more convenient for last-mile transit than buses.
https://sf.curbed.com/2018/6/18/17464616/bay-area-subway-tra...
Electric self driving cars are the future of the US.
In terms of keyword ranking, from what I've observed the only think taking into account things such as keywords with respect to ranking is the readers voting and flagging.
Maybe everyone that is upvoting it can read it b/c they subscribe to the news site, and maybe your comment shows your existing bias about your thoughts on other readers of this discussion forum?
Because Uber's insane valuation is largely predicated on them being able to pivot out of expensive human "driver-partners" into cheap and compliant robots, and any sort of public acknowledgement that this is not going to happen would tank their stock.
A company which can achieve level 5 automation and has the capability to roll it out worldwide is worth a hell of a lot more than $60 billion.
I think the broadest interpretation is that autonomous driving was one of a few irons Uber has had in the fire to achieve profitability. The other irons include Uber Eats, Jump, and lately, grocery delivery. The miracle hasn't happened yet!
The developing narrative seems to be that self-driving was a longshot that could have paid off hugely, had it worked out, but it's been plagued with problems and it just seems very unlikely to pay off now. But still not zero, since no one else has shipped anything.
Investors probably valued it like (0.7% chance of success * 1 trillion upside = 7 billion).
This could increase city trafic capacity by an order of magnitude in congested central zones and suplement public transport in suburban, dispersed areas that cannot be properly served by buses.
And it could be done today, with simple algorithms: stay in your lane, coordinate with other vehicles and the trafic lights, and when anything that looks like an obstacle is in front, brake.
As for the density of such infrastructure, you don't need it on every street, just within a short walking distance to be practical, a few hundred meters maybe. Once you can reach the hail point, it can take you non-stop to any other in the area served, so something like 30 Km/h top speed can go a long way.
We already have trams.
So by using advanced self-driving concepts (but not yet full self-driving alongside human drivers), you can repurpose much of the existing road space for personal public transit, without creating completely new infrastructure.
https://www.theverge.com/2018/9/17/17859112/self-driving-car...
Is that what you mean?
The problem is, once you ask anything of government infrastructure, it may take forever. I could maybe see it eventually happening if these companies lobby heavily for toll lanes, where autonomous zero-passenger vehicles pay the bulk of the fees.
https://www.youtube.com/watch?v=iaSaWfw07Sw
If you used battery electric cars would likely be a lot cheaper to build.
Speaking of electric, in Taiwan they have occasional dedicated paths for scooters. With the advent of battery power bicycles grade separated bike lanes seem more useful. Because battery power eliminates the hill problem.
I actually found that looking for a years old article about Google's system accounting for bicyclists making hand signals, but there you go.
If the worker indicates you to go into the opposing lane, how long should the car remain there? How do you tell the construction part is done and you should return to normal self-driving logic? Sometimes there are cones, but sometimes not. American workers do often have Slow / Stop signs, giving a clear indication of when to move at least.
If you are going down a street slowly, and someone is wearing a hi-vis vest for their own safety, and they point at something, is your car going to follow that pedestrian's instructions?
Deep learning alone has long since been considered a faulty approach (see ALVINN), since it could produce a car that can sort of drive, but nobody knows how, and when a road rule changes you have to retrain it from scratch.
Edit: Of course I meant Uber. Doh!
While I'm sure they have deep learning somewhere in the stack, the core of their algorithm appears to be based on more traditional techniques, with kalman filters being the core.
Interesting! Can you please share a source so that we can learn more?
From the point they got identified as IP thieves it became hard to believe they'd get anywhere.
That’s really interesting. Does anyone have a link to where this was stated?
The settlement is discussed for example here, but no mention is made of Uber having to disclose future developments: https://9to5google.com/2018/02/09/alphabet-waymo-uber-trial-...
I don’t remember how exactly they were going to enforce things though, which is why I asked.
Somehow they felt that they would simply "be the world's transportation" (their stated goal) based on a completely unproven technology.
> Numerous ATG employees have privately said that nobody else at the company appeared to pay a price for the blatant mistakes that preceded the deadly collision and the fact that the units management ignored clear warning signs. For instance, a manager directly warned Meyhofer and other leaders before the collision that the unit's robotaxi prototypes were dangerous-one of them had swerved onto a Pittsburgh sidewalk and continued driving — and that the company did not terminate human backup drivers even if they repeatedly screwed up.
> In mid-2018, several months after the collision, a departing senior manager of ATG sent Khosrowshahi a 35-slide presentation outlining problems in the unit, according to a copy of the presentation that circulated across the company at the time and that The Information viewed. The document alleged that months after the collision the unit still had "inadequate" safety and testing procedures and a propensity for "theater" — celebrating what the team had designed "when we didn't yet have substance" in the form of a working prototype.
> This manager also told Khosrowshahi that the unit's software engineering team was populated by "university researchers" — a reference to the CMU lab where many Uber engineers, including Meyhofer, came from who didn't have much experience in commercializing technology for the real world.
> The former manager declined to discuss the presentation or their experience at the company with The Information. Uber has significantly upgraded its safety practices in recent years, according to Meyhofer and other employees.
> In the same time period, Uber's vehicles had also been having what the company refers to as a bad experience — such as a sudden jerk or a potentially dangerous movement — every one-third of a mile on average. The company's leaders had hoped the figure would fall to just one bad experience every 10 miles by fall last year, said a person with direct knowledge of the goal. One person who worked on the effort told The Information they felt the prototypes were better than the data indicated.
> Uber couldn't even get the prototype to drive a one-mile stretch between the unit's two Pittsburgh offices, with the goal of shuttling employees back and forth. Software leaders said they gave up trying because automating the route for internal use wouldn't help Uber develop software that it could apply to a broader array of routes. But other managers said the failure sent a message that Uber just couldn't pull it off.
> For his part, Meyhofer said road testing helped the unit's broader mission, including giving the team ideas on how to improve the way it configures the car's laser sensors. "Every time we do any kind of development, it shows us a lot of things we can do that will get us even better results" in the next test, he said.
> Meanwhile, the company is focusing on developing new autonomous driving software to replace the version now in its prototypes, which includes code academic researchers wrote about a decade ago. The new software, rNA, is nearly ready for testing on public roads. Meyhofer and other current and former employees said INA-which took longer to develop than originally anticipated-was a necessary upgrade to allow the company to incorporate new machine learning-related techniques that help vehicles better understand the world around them.
Self-driving tech should be focused on driver enhancement and not replacement. This is an incremental evolution area until all the 100 edge cases are 100%.
The self-driving enhancements they could give to existing Uber drivers to help them stay safe over long driving periods would be huge. Drive that flywheel (easier driving leads to more drivers, cheaper trips, more riders, etc.).
Let drivers multitask under safe scenarios in under 10mph traffic jams, open areas, queues, etc.
Build up those enhancements and you have a formidable, cheaper driver fleet that people would use perhaps v. buying a car.
Just saying it should happen in stages.
I feel as though paywalled links should be heavily downweighted as most of the discussion will uninformed.
To answer your question, Outline.com doesn't seem to work here.
Incredibly frustrating to read comments like this. We can't rely on ad funded and corporate funded journalism forever. Downweighting actual attempts to fund real journalism is completely misguided.
By HN design, if such sites aren't useful to the HN community, they wouldn't be regularly upvoted. However counter to that, every week you see an article from wsj.com, theinformation.com or bloomberg.com hit the front page, precisely because despite having paywalls they continue to produce good content. Given the amount of engineers sucking six figure salaries on this site, I'd say a good percentage of them who might value actual news may pay for a subscription, which is why these sites continue to be posted here. I wouldn't want to see wsj.com downweighted below buzzfeed.com just because buzzfeed is free.
The ones who choose to complain, rather than just abstaining, are the problem. They could simply decide to not participate and HN wouldn't be any worse off. However the complaining that news organizations must (1) offer all news for free (2) don't use ad networks and (3) not be mouthpieces corporate PR is not going to accomplish anything.
Would some kind of HN-centric wire service help?
Self-driving tech should be focused on driver enhancement and big replacement. This is an incremental evolution area.
The self-driving enhancements they could give to existing Uber drivers to help them stay safe over long driving periods would be huge. Drive that flywheel.
Let drivers multitask under safe scenarios in under 10mph traffic jams, open areas.
Build up those enhancements and you have a formidable, cheaper driver fleet that people would use v. buying a car.
Boil the ocean and you’ll be useful merely for your learnings to others.
The theoretical arguments about how self-driving cars are better than human drivers remain theoretical.
Interesting, can you please share a source?
Would nobody sell them such tech or do they think anyone who solves this problem will put Uber out of business?
It's a moonshot, and the first to get it right will make far more than $2.5B in revenue.
Edit: I realize that level 4 doesn't ask you to take over, so I've edited my comments to reflect that.
Edit: I realize that level 4 doesn't ask you to take over, so I've edited my comments to reflect that.
ex all the horror stories of distracted Tesla drivers using autopilot like full autonomous driving today.
And, yes, people imagine scenarios like remote drivers taking over. That seems a very difficult problem--especially if the situation is such that the autopilot is confused.
I also have very little confidence in the AI's ability to stop at road construction, make its way over temporary embankments and diversions placed for construction, navigating in inclement weather or adverse traction situations like a muddy drive, etc. I suppose the real goal here is for Wal-Mart and Amazon to move freight on the Interstates to begin with but I don't even see that happening any time soon. Your average Wal-Mart loading bay and parking areas are likely much more than any current AI can handle anyhow. How does the AI know the person in the mart cart isn't going to just jut out in front at the last second? Can't exactly make eye contact.
"...one of the most advanced robots in the galaxy that can do billions of computations a second and you ask me to do the most mundane tasks... go open the door Marvin. Go pilot the ship Marvin. Go fix the reactor Marvin ... Go drive my Tesla Marvin ... I think you ought to know I'm feeling very depressed..."
It's mostly a problem of sensor fusion. Once your self driving car has a good view and solid understanding of the surrounding environment, the actual driving algorithm is reasonable simple. All you really need is really solid software engineering to implement it with minimal bugs.
The problem is getting that understanding. You can't have objects pop in and out of existence. You can't have computer vision misclassifying objects or just not seeing them.
It might just require better sensors, or more types of sensors. Or we might need to work on interpreting the data and understanding the actual situation.
Most likely it's a combination of both.
https://www.sae.org/news/2019/01/sae-updates-j3016-automated...
Arguing about what to call it is boring. Is Tesla close to having a system that can drive without attention, decide it isn't sure what to do and safely request driver attention? Or does it just have the bit that can successfully navigate a lot of situations?
Edit: if I didn’t make it clear, Tesla vehicles safely hand over responsibility to the driver when path planning confidence has been lost. If you don’t take over, the vehicle comes to a stop with the hazards on. If you haven’t, I recommend taking a free test drive yourself to better understand the system constraints and UX.
Much of the distinction between level 2 and level 3 is that the vehicle reliably hands over control. Successfully navigating situations doesn't provide much information about how close that capability might be.
You've been lucky then, not to experience phantom breaking, cutting off other cars when changing lanes or fire trucks stopped partially in your lane.
It's good that the Tesla system stops making inputs when it doesn't know what to do. That's different than the Tesla system having full control and safely notifying the occupant that they need to start driving.
It seems like level 3 is more dangerous than a 100% manual car.
Even high-end level 2 seems dangerous.
Until we get self driving cars to level 4, were they don't require the driver to play attention, I don't think self driving cars should be on the road.
How exactly would you characterize the comparison?
Modern vehicles (ICE or EV) are vastly more sophisticated than "putting together metal on an assembly line" -- they're modern marvels of all the physical sciences: chemistry, thermodynamics, materials science, electronics, mechanics, aerodynamics, wireless connectivity, reliability, safety, real-time controls, computer vision, sensing, and on & on & on. Vehicles are vastly more "technological" than most mobile phone apps, especially one like Uber.
Pure software companies may (rightfully) fetch higher valuations, but that's simply due to the nature of their economics: higher margins, pricing power (oligopoly?), and negligible incremental unit costs.
But so far, Tesla is dead last in self-driving cars. They some cool driver-aid systems, but their self-driving tech is still being written and rewritten, and didn't even begin any tests on the roads.