Inside Waymo's Secret World for Training Self-Driving Cars
theatlantic.com
theatlantic.com
If you want to do it right, you start with geometry. The first step is capturing range imagery and grinding it down to a 3D model of the world. This tells you where you physically can go. That's where we were at the DARPA Grand Challenge over a decade ago.
Then comes moving object popout. What out there isn't a stationary obstacle? That comes from processing multiple frames and range rate data from radars. Moving objects have to be tracked.
Only then does object recognition come into play. Moving objects should be classified if possible, and fixed objects of special interest (signs, barricades, etc.) identified. Not all objects will be identified, but that's OK. If it can't be identified and it's moving, you have to assume the worst case - it's vulnerable and it doesn't follow the road rules. This will force a slowdown and careful maneuvering around the unknown object. (See Chris Urmson's SXSW video of their vehicle encountering someone in a powered wheelchair chasing a duck with a broom.)
Predicting the behavior of moving objects is a big part of handling heavy traffic. That's what Google is working hard on now.
Machine learning is a part of this, but not the whole thing. You can't really trust machine learning; it's a statistical method and sometimes it will be totally wrong. You can trust geometry.
If you want to do it wrong, you start out with lane keeping and smart cruise control, tack on recognition of common objects, throw in some machine learning and hope for the best. This produced the Tesla self-crashing car, noted for running into unusual obstacles which partially block the lane and for trusting lane markings way too much. Look closely at the videos from Tesla and Cruise, slowing them down to real time. They speed them up so you can't easily see how bad the object recognition is.
I would also add that using simulations to test (or train) robots is also not new, and has significant limitations. Robotics has been using simulation since before the Grand Challenge, but what we find is that simulations make simplifying assumptions that don't bear out in real life: simulated sensors are more idealized than in real life, object recognition is idealized, actuators are idealized. But that makes it hard to apply any learnings to the real world, where sensors hallucinate and actuators lie all the time.
And yeah, we _could_ be fuzzing the simulation and adding failure modes, but not many people do. It's hard to simulate these kinds of defects in a usefully realistic way.
The key thing to understand is that your simulator is not a fixed product provided by a third-party. It grows organically with your codebase. So any time you encounter a bug in the real word, you have a three step process:
- Create a test case to reproduce your bug in the simulation. This may involve updating your simulator to simulate the new physical properties we care about (such as new failure modes or new types of the sensor noise)
- Fix your code, and make sure your test case passes
- Add the test case to automatic testing suite and keep it there *forever*. this is the most important part.
Unfortunately, many academic courses use an opposite approach: they use pre-made simulators, and do not modify them at all. This gives students the wrong expectation that you can find a simulator which can just simulate everything perfectly out-of-the-box; when this (predictably) fails, they become disappointed in the whole simulation idea.
Thank you! Can I ask why the self-driving car programs like the 3-semester Udacity one deliberately focus on going about it the wrong way ? Why spend 1 whole semester on lane detection using trivia like OpenCV, 1 whole semester on object classification using CNNs with Keras, spending so much time on combining lidar & radar data using Kalman filters for moving object tracking .... surely all of these are peripheral issues and the wrong way to go about the business. Then why ? Is it all one big clueless scam ?
To be fair, I am enjoying the course very much as a student. But it's becoming clear from talking to colleagues in the autonomous car industry that what is being taught is not the real deal. This isn't what happens in production, so to speak.
I'm not sure learning about natural language processing or image classification would be all that different--although those problems are arguably better understood and more bounded.
Even a site reliability engineering course isn't really going to give you deep insights into what, say, Google does on a day-to-day basis.
I think this really gets to the core of the issue. There aren't many courses that will accurately reflect the actual state of the art in any field. Sometimes, graduate courses with a small scope and a relevant professor taught in person can partially address the state of the art, but an online course about "how to self driving car" is not that.
By the end, you'll have a few toy projects and a very basic ability to converse about or study more deeply these topics. Maybe they should have spent a little more time on maps and SLAM instead of other stuff, but hey, it's their first version anyway; I'm sure they'll iterate (as they already have somewhat in response to feedback). They did kind of admit in the CV introduction (lane finding) that this was kind of a super simple, satisfying project to start with. I feel like they often touch on topics acknowledging they have more of a historical relevance and are no longer state-of-the-art tools.
(Edit: I should also note: while I have criticisms of the program, overall I'm very happy that I took it. I'm sure some people are better at seeking out specifically what information is relevant to dig deeper into, but I really appreciate the broad overview kind of pointing me in the right direction. Since starting the program I've amassed quite a list of resources to dig deeper into, and hope to eventually find work in this or related field)
Self-driving is something many people in software haven't seen much - software that isn't feature-oriented. As someone from Google pointed out, the output is just two numbers.
As you suggest, for at least some classes of problems, we'll find that this approach will asymptotically approach some level of "good" that turns out to not actually be good enough. And the ways to get beyond that will be hard and may not be obvious.
If the AI is confused by a roundabout as described I wonder how they handle that or if this is something that only fools me because I am only using visual sensors.
It would also be interesting to know whether Waymo factors in other cars' behavior. If everyone is going 80 while the speed limit is 65, will it factor that in and go 80 or keep with the speed limit?
The speed limit is kind of interesting, occassionally see someone go the speed limit on I-5 but it's hard to maintain because there's always an eighteen wheeler going slower than that, and then it hits a hill, and then there's a cascade reaction where the slow lane hits 45-50 for an extended period but the fast lane is going 75-95, though just being patient and waiting in the slow lane isn't a big deal for a autodriving car, or 280 between San Francisco and San Jose with sections where it appears almost everyone is going 85 miles per hour (not sure why CHP doesn't just camp out there). I see Teslas in both these places but I've never seen a Waymo.
No, that's illegal to drive over the speed limit.
Now add in a poorly marked road dense traffic and construction. Other drivers (as a ground) don't really have the spare attention to allocate to the fallout of someone who's slowing down.
edit: This is backed up by traffic studies When everyone is driving the same speed things are safer because there's less "friction" (I think that's the correct term) between traffic elements.
Too slow was equal to too fast during drivers ed.
We had a trailer with a simulator screen and a projector, and about 10 or so fake wheels, brakes and accelerator. He would stop the reel and quiz us about the scene (kid on bike about to cross road; kid hidden by truck. ) The instructor had a print out of our actions. Pretty sophisticated for the mid '70s.
How else do they ever want to work in egypt, or around the arc de trioumpe(sp?) in paris?
As the article points out, Google has far more autonomous driving miles than everybody else put together. They're on the hard cases, too; not just driving on freeways.
Tesla being able to record sensor/map data from human drivers driving through snow will help them mitigate this. On the other hand, electric batteries have less range in cold whether, so I imagine that in the near term electric cars will be less popular in snowy markets versus, say, California.
https://cleantechnica.com/2017/06/23/tesla-electric-cars-con...
I was surprised to read in the article that they were flummoxed by roundabouts, of all things. And that they were apparently surprised to discover that multi-lane roundabouts even exist at all. Have none of these engineers ever gone on holiday to Europe? If they think a multi-lane roundabout is an advanced case, what will they make of this?
https://www.wired.com/2016/08/brilliant-sorcery-englands-7-c...
Overall this article does give hints though that autonomous vehicles are truly much further away than anyone wants to admit. The story about being flummoxed by multilane roundabouts is depressing. If they didn't know such things existed then they were incredibly sloppy in their data gathering. If they truly thought scaling from a simple roundabout to a multilane one would be trivial then their staff don't have the right mindset for this problem space.
Also note that all the testing is happening in flat, desert landscapes where there are no weather or lighting challenges. Their pictured model residential street only has stubs of driveways. No houses, no trees. I'm sure they know these are gaps but I worry they underestimate the challenges of adapting to entirely different driving environments. Especially when machine learning is involved and you're simulating 99% of your mileage...
Looking forward to checking back in 2040 though.
I'm wondering why you say that? Wouldn't the more relevant factor be the type of miles, regardless of whether they are simulated or real? For example, 1 mile of simulated "Interesting Scenarios" (Duck on Street, Bicyclist going wrong way, pedestrian running onto road) is likely worth 100,000 miles of normal Freeway driving.
To get close you have to pay people to send their kid running between parked cars into the street at just the right moment. Nobody sane will put a real kids into that situation, but the simulation can (should!) have thousands of different variations. The simulation can afford to run all of those situations regularly. Simulation can ensure it is raining when you need it.
You need miles of real driving too, but they are too expensive and unpredictable to rely on for more than a final check.
Don't underestimate the bizarre engineering of infrastructure in Texas. Every road and intersection is bespoke, unlike any other you have previously experienced. It is quite difficult to navigate Austin as a human driver, I suspect it makes for a good testing ground for Waymo.
> “GM, Ford, Toyota, and others are saying ‘Let me reduce the number of crashes and fatalities and increase safety for the mass market.’ Their target is totally different,” Peng said. “We need to think about the millions of vehicles, not just a few thousand.”
I wonder if an easier task to reduce fatalities is to target impaired driving and develop a detection mechanism that tries to identify whether the driver is impaired. It seems like an easier and more localized problem than building a fully autonomous vehicle.
EDIT: Apple maps has up to date satellite imagery https://maps.apple.com/?q=37.3718,-120.5749&t=k
Apparently it's been doing this for years, and in different platforms.
This is so far from being my area of expertise. Just one observer's thoughts/questions. Insane how far we have come so quickly, at any rate.
That is because much of the machine learning research is coming from industry, particularly the mega-corps. And for extremely important disruptive events such as AI, they are willing to drop $$ into collecting a stable of research scientists to not fall behind in this space. So while the AI winter was driven by a drop in academic and government spending on research, I think our present situation is far less likely to result in a similar drying up.
Start ups could get hit, however.
Because the company is a king, and the employees are the property of the crown?
I found the migration of experts from CMU to Uber fascinating — for two reasons: 1. The geographical proximity in which the acquisition occurred (i.e. a rust belt town with an awesome CS school and a well known Silicon Valley startup on location). 2. The moving from the academy to the private sector of said professors/researchers happened in a way I hadn't seen before. I mean lots of academics work within industry at some point, but this move seemed to carry more weight in the media — maybe because of the institution involved.
Anyway, with that said, lots of the articles and speculation I read about it, for better or worse, painted the acquisition of academics by Uber as a bit unfair. I.e. Uber could and did pay way more than CMU and provided super interesting problems maybe outside the realm of what a professor normally faces in their research, I don't know. It was the "unfair" element portrayed in articles/opinion pieces I based my wording off of. I am in no way suggesting these profs and researchers did anything wrong or did anything but make the right choice for themselves, something we all have to do.
I digress. Word choice noted, problematic history and other associations noted and hopefully I cleared up my choice a bit.
- TTS and speech synthesis have lots of uses in Android (phones, Wear, Auto, etc.)
- Object recognition is very useful for photo search
- Face recognition is also useful for photo search (if you tag people in Google Photos)
- Neural machine translation provides better translation accuracy (for languages which we have enough training data for)
It's possible Google will cut back on some more speculative ML research efforts, but certainly not those four, I would think.
As you've mentioned, object detection (which is a considerable portion of CV) is marketable. Meta learning (Vizier, "Learning to learn with gradient descent with gradient descent") are incredibly valuable from a business perspective. Model compression is important for neural networks on mobile/embedded devices, etc.
One of the very interesting things about ML being such a corporation driven field is that it's incredible how quickly relatively recent research makes its way into products.
I think this is just true for really anything that falls purely within the domain of the computational sciences. For a ground breaking drug to hit the market, it takes 10 years of FDA trials, for high speed trains it requires large teams of construction workers only after years of safety work.
In the world of software as it stands today, deploying a new model for something can be virtually instant in some cases, and even in the worst cases, it isn't taking people years to improve a small attribute of their product.
How much money does Google Photos make? Last time I used it, there were no ads in it.
Isn't speech recognition nearly at human levels of comprehension, at least for US English and outside of highly technical jargon? Speech patterns change slowly, so would Android continue to make money if investment in speech R&D were cut back? My guess is yes, it'd be fine.
Does Google Translate make money in some specific, measurable way? I don't mean "well it's neat to have translation links in web search", I mean, would people cut back on the commercial queries that are Google's main source of income if Translate quality stopped improving? I doubt it.
Much of their products are like this. They're products but they aren't businesses. It's very easy to lose sight of the basics of business when inside the Google bubble. The things you mention are not providing business value in the conventional sense of the term because they aren't yielding independent profits. So they're all vulnerable.
That said, vulnerable to what? I think Google is at risk of losing a lot of trust around web search and political and informational search more generally, but commercial queries are probably pretty much impregnable.
Google Photos is, in a sense, a big ad with prominent calls to action for selling storage space.
> Isn't speech recognition nearly at human levels of comprehension, at least for US English and outside of highly technical jargon?
Recognition, maybe (though I think no). Applying recognized text to get the desired outcome, not even close.
> Does Google Translate make money in some specific, measurable way?
It powers a paid API, so, yes, it makes money in a specific, direct, measurable way.
FYI: there is no such thing as a "neutral accent" - unless you mean neutral to your locale. I could describe an object as having "neutral temperature", but you'd need to know how hot/cold its environment is before it makes sense.
It's true that this accent generally lacks the strong regionalisms that some other US accents do. But it's mostly still an outgrowth of a general region even if its use is cultivated more widely.
Recognition of technical jargon is an easier problem than recognition common speech. Jargons are far more regular and have limited lexical and idiomatic variation, while common speech has a huge number of variations and is constantly innovating new usages. This is evidenced by the fact that the earliest speech recognition systems were only useful in highly restricted domains (medical transcription, phone support tree navigation).
It will be either one of those two, if it doesn't happen in the next 20 years I would bet against it happening at all until AGI is a reality, assuming it will ever happen in the first place.
But if it is possible with an incremental change to present day technology I fully expect it to happen and likely a lot faster than those 20 years. The economic incentives are too large to ignore and there are many players wealthy enough to finance such an incremental step.
It's not sexy and doesn't fulfill autonomous taxis everywhere fantasies but for the vast majority of people who don't live in or near urban cores it would likely be a significant quality of life improvement.
It seems a best of both worlds solution would be a new type of job: Data gathering driver. It would be basically like the Google Streetview cars, but with much more sensors and input from the driver. These drivers would be assigned neighborhoods to drive through and note anything unusual or noteworthy. Stop sign is down because of last night's storm? Driver notes this and now all Waymo cars know. Street closed due to construction? All Waymo cars will avoid that street. This type of data requires intelligence to analyze and understand and it would be fed into the Waymo system for all self driving cars to benefit from it. There could be millions of self driving cars and a few thousands of these drivers feeding daily updated data to the system. You could even increase the frequency of the drivers depending on weather conditions. Maybe if there's a storm, the drivers keep driving in a loop hourly instead of once daily.
It doesn't have to be perfect. Just very good and improving. Some time ago google shared a gif of a wheelchair chasing a duck in the middle of the road. The car didn't understand it, so it just stopped. Good enough for me.
Obviously, they have a lot of work ahead of them, but don't be so pessimistic. Most people drive shitty (myself included), we aren't impossible to improve upon.
Stopping might be a sensible safety protocol in some situations, but it isn't in others (not to mention the situations where the car may stop too late because it doesn't actually recognise that a novel scenario is about to occur).
So if you want Level 5 driving, and not just a very impressive demo which is safe provided a human watches it attentively enough to take emergency measures and is able to override it when it decides it can't process a situation well enough to proceed, you need the AI to be pretty damn close to perfect in its judgement of how to react to a huge number of edge cases.
So, self driving cars can simply be very cautious without seeming to.
EX: I am sure the car had no idea what this was: https://youtu.be/Uj-rK8V-rik?t=26m11s but as long as it can tell it's bigger than a bread box and so it should not to hit it that's enough.
Similarly, instantaneous harsh braking might help an AI save the small child it didn't anticipate might chase the football that flew past moments earlier, but a human capable of grasping that footballs are associated with pedestrians making rash decisions might have braked early and gently enough to not get bumped by the car behind. (If they didn't, they might find their late reaction blamed for the accident and possibly even get prosecuted for driving without due care and attention).
The UK requires every learner driver to sit an exam consisting of identifying CGI "developing hazards" where they're scored on ability to rapidly identify stuff that might happen before they're allowed to do the full driving test. I'm sure a key focus of the teams the article discusses is teaching AI similar cases like gently slowing in the event of a football-shaped object moving near the road (which is likely far from the most difficult or obscure novelty to teach an AI to handle) but the problem space of novelties humans handle by understanding what things might be and how/if they are likely to move isn't small or one there's good reason to believe plays to AI's strengths
(Meta: not sure why you're being downvoted, your contribution seems constructive and on topic to me)
Which IMO is what's missing from the debate, unusual events are in terms of ~10+ million miles of training data before these things are in production. They are clearly out there, but I doubt people are going to react well to say someone falling from an overpass onto the road very well either. So, it's that narrow band of really odd but something a person would respond correctly to that's the 'problem'.
PS: Of course the bag might relate to a bug which are likely. But, IMO that's a completely different topic.
Opting to drive around a (stationary, visible from a distance) bag in an unpredictable manner is literally how Waymo's first "at fault" accident occurred...
The point is that a human has a concept of a "ball" linked to the concept of "children play football" and an understanding that if one sees the former, one should be prepared for the latter to bursts onto the road from behind the partially-obscured roadside. Appropriate action probably involves easing off the accelerator and lightly tapping the brake so the car behind gets a hint that you might have to stop suddenly on if a child emerges from behind a bush. An autonomous car which fails to anticipate even though it's lightning fast at slamming the brakes on is going to get rear ended a lot more.
The neural network of a self driving car might be able to classify small coloured spheres in the vicinity of the roadway as balls, and the AI will certainly have been taught the concept of a human-shaped obstacle moving across the roadway being a "need to stop" situation, but is unlikely to "learn" the association between the two through a few tens of million miles of regular driving, because only a very small proportion of "need to stop" events involve balls (and only a very small number of sightings of spheres moving in the vicinity of the roadway result in "need to stop" events). Of course, you can hard code a machine to respond to ball-shaped objects moving near roads by slowing down and you can construct a huge number of artificial test scenarios involving balls to teach the AI the association between balls and small children, but either of these options involves engineers envisaging the low frequency hazard and teaching it enough permutations of the sensory input for that hazard for it to be able to anticipate it (and there's a balance to be struck, because nobody wants a paranoid AI which drives through the city braking every time it sees something its neural network identifies as a pedestrian or the front of a parked car protruding from a driveway) Suffice to say, we take for granted our ability to know how to react to things like children chasing balls, staggering 4am drunks, tiny puddles the car in front just drove through versus a raging torrents of water through the usually safely navigable ford, sandbags versus shopping bags, vehicles laden down with loads which look like things which are not vehicles, and people frantically gesturing to stop.
As to the ball case, it does not need to stop just slow if it's possible for a kid to run out fast enough to be a problem. aka on residential streets when it should not be going fast in the first place not a highway where there are no major obstructions. So again, what your describing is in the 'to be dealt with' pile.
I don't mean it's working right now, just they are going to hold off on mass production until it's working.
There wasn't anything unpredictable about the behavior of Google SDC. From the official statement: https://www.engadget.com/2016/02/29/google-self-driving-car-...
"Our car had detected the approaching bus, but predicted that it would yield to us because we were ahead of it.
Our test driver, who had been watching the bus in the mirror, also expected the bus to slow or stop. And we can imagine the bus driver assumed we were going to stay put. Unfortunately, all these assumptions led us to the same spot in the lane at the same time. This type of misunderstanding happens between human drivers on the road every day."
Anecdotally, Uber has been gathering lots of road quality data from its drivers for years. Not sure if it still holds, but they were recording and uploading accelerometer data for a while.
[0] https://www.reuters.com/article/us-lyft-waymo-collaboration-...
[1] https://www.whitehouse.gov/the-press-office/2017/04/06/presi...
The generated geometry doesn't seem to include the Bike that quickly passes behind the car at the intersection ...
The wireframe is an artificial simulation that is inspired by the same scenario.
Isn’t that 1.12~ per 1000mi?
To my surprise, railroad travel is more fatal at 29.0 deaths per billion miles traveled. [2]
[1] https://en.wikipedia.org/wiki/Transportation_safety_in_the_U...
[2] http://www.caranddriver.com/features/safety-in-numbers-chart...
https://www.rita.dot.gov/bts/sites/rita.dot.gov.bts/files/pu...
you can see that 95% of railroad fatalities are classified as either "Trespassers" or "Highway–rail grade crossing". In both of those cases it seems likely that the victims are generally not actually passengers on the train, but rather are people doing something obviously dangerous near railroad tracks.
(Also the denominator wasn't clear for the quoted statistic: is it per passenger-mile, or per vehicle-mile? In any case here is a sensible denominator: https://www.rita.dot.gov/bts/sites/rita.dot.gov.bts/files/pu...)
I think the equivalent for a human driver would be a situation where they're unsure what to do, so they freeze up, or proceed extremely cautiously. I'm not prone to freezing up, but I've definitely encountered situations where I'm uncertain to the point that I proceed with the utmost of caution.
Human disengagement isn't easy to define. People don't just say "well this is too confusing, I'm just gonna turn control over to someone else". To define disengagement you need to define a performance minimum. An accident is about the lowest floor you can set to I used that. I would personally say that if you hit the brake on an on ramp while traveling slower than the highway traffic you're "disengaged." People and AI generally travel as fast as they can while maintaining a sufficient safety factor (e.g. people generally don't tailgate each other at triple digit speed and drive slower in snow). If someone's personal safety factor/comfort level requires several minutes of everyone else honking while they wait for a space fit for a semi-truck before pulling their Prius into traffic that may be acceptable to them but if an AI did that I think most people would say that's a failure even if the AI didn't turn control over to a human.
Tesla just enabled in the last few months the option for people to upload data to them. Determining which data to upload is its own problem as well.
Did you mean Tesla never captured data before that? If yes then I disagree. Sorry.
More here - https://electrek.co/2017/07/12/tesla-global-fleet-electric-m...
Tesla just put out there an opt-in upload of video data 3 months ago. And even then it is not clear how useful that data is. Especially since a v2 of the autopilot hardware was recently announced.
https://electrek.co/2017/05/06/tesla-data-sharing-policy-col...
Tesla has an enormous amount of data at their disposal. The recent change concerned video footage from the onboard cameras.
From what I've learned, lidar and radar are typically used for object detection (i.e., avoiding other cars; perhaps augmented with camera data), while cameras are used for things like lane detection and traffic sign detection. If Tesla is trying to solve the same hard problems as everyone else, using weaker equipment, I wouldn't bet on them having a lot of success.
Also, kind of the point of this article is that Waymo has a test area where they can get real data along with the ground truth. That seems much more valuable than unlabeled/random data.
1) They believe no one can quite catch up before they can launch the technology. Since they know several competitors have huge resources and brilliant people, it means they are quite close to launch.
2) These tools have become open secret within the industry, so no harm is done to their competitive position by revealing it to the public, only good PR to be gained perhaps to attract more bright engineers.
I suspect 2) is more likely since several key players have moved around so much in the past couple of years. Relatively high-level knowledge about how autonomous vehicles are being developed at Waymo might have become well-known within the industry by now.
Edited: of course Murphy's law; thanks for pointing it out!
In my medium-sized metro area, I can recall at least a half-dozen deaths of cyclists this year caused by humans making similar errors. It doesn't take much for software to be better than that.
I wish them the best of luck in their continued progress! I'll stick to easier stuff :)
The same goes for dense fog, flooding and other situations where normal driving is no longer possible and/or responsible.