14 ms·
Waymo’s autonomous vehicles are driving 25,000 miles every day
- trhway 8y ago5B miles in simulation. I wonder do they source the situations for simulation from Youtube from crazy/fail/accident driving videos.
- y4mi 8y agoI can't imagine these poor video streams are useful inputs for the algorithm... Maybe simulations are when the human is controlling the vehicle and the algorithm is only allowed to .. simulate what it's actions would be if it were in control?
- toast0 8y agoIf they've been recording everything, they can run the previous inputs on new algorithms and check the results, and also simulate sensor failures and what not. Running the system on recorded inputs to see how it compares with human output is good too.
- Fricken 8y agoWhen they encounter interesting situations while testing on real roads, they then take that situation over to the simulator and train their algorithms on 1000s of subtle variations of the situation, which Waymo calls 'fuzzing'. Simulating effectively is non-trivial. In spite of testing on a comparable scale to Waymo, Uber made very little progress in 2 years because they couldn't get their simulator to work right.
- ceejayoz 8y ago> I can't imagine these poor video streams are useful inputs for the algorithm... Store a rolling last X minutes of full sensor data into memory, then dump it to disk if the supervising driver hits the "that was a fuck-up" button.
- repsilat 8y agoThey don't use the video data in the simulations, they use the processed "Where everything is" stream. That makes it more robust to sensor tech upgrades, but less to errors to do with bad sensing or sensor failure. Computationally cheaper and simpler, no doubt, too.
- kickopotomus 8y agoThe source the simulations from sections of road that their cars had trouble with. https://www.theatlantic.com/technology/archive/2017/08/inside-waymos-secret-testing-and-simulation-facilities/537648/ https://www.theatlantic.com/technology/archive/2017/08/insid...
- TomK32 8y agoRussian crazy/fail/accident video trump every US youtube video.
- tim333 8y agoThat would be entertaining to watch. Possibly also useful to check the cars don't freak in weird situations. They kind of had a real one in the Chandler crash https://www.youtube.com/watch?v=vxqBS2-4puw https://www.youtube.com/watch?v=vxqBS2-4puw
- Fricken 8y agoWaymo has tested in 25 cities, they aren't currently in 25 cities. Of the 600 Pacifica's they have, only ~160 are deployed. So to do 25,000 miles per day, that's an average of 156 miles per vehicle.
- blacksmith_tb 8y agoThat does seem like a lot, I suppose they theoretically could be driving 24/7/365 (the wear and tear would be serious, you'd think). The idea of moving people to public transit sounds good, but I wonder if there's enough spare capacity in most transit systems to serve lots of new riders...
- vkou 8y agoThere is spare capacity in most transit systems to accommodate new riders, just not at the SLA that prospective new riders are accustomed to. If your suburban bus stop is a 10 minute walk from your home, and the bus comes by once every half-hour, and then you have to transfer to an every 20-minute bus, that is sufficiently offset to make the transfer take time... You aren't going to have a good time on your commute.
- nostrademons 8y agoThat seems pretty reasonable - 8 hours/day at 20mph, 5 hours/day at 30mph, and 2.5 hours/day at 60mph. Given a mix of highway, rural, and city driving, 156 miles/vehicle sounds about right.
- JoshuaEddy 8y agoAnd only 6.5 MPH if they're running 24 hours/day
- dzdt 8y agoHuman drivers average about 1 fatality per 100 million vehicle miles traveled. [1] So we are still a long ways from being able to access safety in comparison to humans. [1] https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812451 https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/...
- hcnews 8y agoNumber of fatalities per million miles driven is only one of the metrics. Hours spent per lifetime driving is arguably a more important metric. I can't wait for everything (cities, highways etc.) to be redesigned for self-driving cars but looks like its not going happen in my life time :(
- pvarangot 8y agoIf you have to sell self driving cars as "more people will die but some people will have more free time" it's not going to be as easy as if they are unarguably safer for everyone.
- xyzzyz 8y agoI don't think that's true. That's exactly how we were sold on cars in the first place -- you're more likely to die when you drive than when you walk, but you can get places faster to save time. Same can be said about e.g. caffeine. Humans are notoriously bad at estimating and considering tail risks.
- Smoofer 8y agoCare to expand on what you mean by risk vs. return of caffeine?
- natrik 8y agoPeople drink coffee (consume caffeine) because they wish to be more alert/are sleepy. This is a consequence of not getting enough sleep. Lack of sleep has many documented negative effects, but people trade sleep (health) for productivity/etc. via caffeine intake.
- RestlessMind 8y agoAs an HNer, I find what Waymo (and other autonomous vehicle companies) are doing is simply fabulous. Technology improvement is a big impetus for human progress. Looking back over the past century, inventions like transistor, rockets, internet, satellites, nuclear power and aviation significantly improved our quality of life. Autonomous vehicles belong to the same league. I am very happy that a non trivial amount of resources are devoted to them and it's not just a next SV fad like some ICO / Juicero / photo sharing app. Rooting for the success of this technology.
- mrlatinos 8y agoThis isn't Reddit, no need to prefix an otherwise normal comment with internet credentials.
- YeGoblynQueenne 8y agoCars destroy the environment (with the infrastructure needed to accommodate them) and are one of the primary drivers of climate change. We need technology to make them obsolete, not to make more of them that can drive themselves (or not).
- eboyjr 8y agoThanks for playing Devil's Advocate. This technology applies to mobile transportation in general and it doesn't necessarily have to effect greenhouse gas emissions. Additionally, I believe the infrastructure for producing cars applies to such a broad category in regards to driving climate change. The manufacturing industry needs to independently solve this problem.
- droopyEyelids 8y ago> ...the company is also working to apply its self-driving system to three other areas, including logistics (so trucking)... This whole 'disenfranchised underclass' thing that America is struggling with will get considerably worse when automation turns all the truck drivers and their families out onto the street.
- RobertoG 8y agoWhen automation arrive to that level, that could happen, but it's far from unavoidable. If automation happens is because it makes the economy more productive. If before happening the economy could sustain all the people, logic tell us that now could do it even more easily. At the end, it's a political decision.
- CydeWeys 8y agoMost people used to be farmers, too, but prohibiting industrialization was not the right call. Driving trucks is a dangerous occupation that's terrible for your health, and often takes you away from your family for long periods of time. In a few generations when it's long gone as an occupation, no one will miss it. Getting from now to then is unfortunately going to be a rough transition for some people, but stopping that transition entirely isn't the right call.
- EthanHeilman 8y ago> Most people used to be farmers, too, but prohibiting industrialization was not the right call. The industrial revolution caused enormous suffering for ex-farmers. Many of them lost a life of autonomy to be forced to move to cities and work in dangerous conditions for starvation wages. It caused significant social instability. https://en.wikipedia.org/wiki/Industrial_Revolution#Standards_of_living https://en.wikipedia.org/wiki/Industrial_Revolution#Standard...
- IshKebab 8y agoAnd I think that is the place it will happen first. At least outside America where most roads are far too complicated for driverless cars. Motorways, on the other hand are relatively simple. And we have service stations that driverless lorries could drive themselves to, to be picked up by human drivers for the final leg. And there's an obvious commercial need. And there's an obvious safety need (who hasn't seen a swerving lorry?)
- tlrobinson 8y agoWaymo is supposedly planning on launching a ride-hailing service. Perhaps they can start giving rides to customers while continuing to test with a live driver (to take over if necessary). The revenue from the ride-hailing service could offset the cost of testing, and allow them to operate indefinitely until full autonomy is ready and legal.
- bob_theslob646 8y agoUnfortunately the liability is too high for them to start taking passengers without thorough. The idea is pretty clever. People paying to be guinea pigs. Sure if people signed off, it would reduce some of the liability but that's still the golden question: who's responsible if something happens?
- ghaff 8y agoIANAL but seems pretty obvious. If I'm just sitting in some company's product and it causes harm to me or someone else, it's 100% on the company. Of course, anyone can be sued for anything but the idea that I could be criminally negligent because an autonomous vehicle I'm sitting in (whether car or train) killed someone seems absurd.
- jacksmith21006 8y agoWaymo has insurance from Munich Re. Do not think liability problem. https://www.insurancejournal.com/news/national/2017/12/20/474893.htm https://www.insurancejournal.com/news/national/2017/12/20/47... Startup Trov, Munich Re to Insure Riders of Waymo's Self-Driving ...
- tlrobinson 8y agoYou can insure just about anything. At some point Waymo will be able to show the risk is low enough that insurance will be economical, if it’s not already.
- bdod6 8y agoride sharing isn't exactly very profitable at the moment. I don't think the revenue would offset the additional engineering/product costs of just running the rideshare service, much less the cost of testing.
- vannevar 8y agoThe number of miles is not a sufficient metric; the question is, how much of the automotive phase space (weather, traffic conditions, terrain, road surface type, time of day, etc) has been explored? Running over the same mile a million times does not add anything to the self-driving model.
- scarmig 8y agoInterestingly, going with that logic disengagements per mile will at some point correlate with the most advanced AV companies, not the least. I guess the response would be to target some SLO budget and, once disengagements get too high, refocus on existing cases to get it down to an acceptable level.
- agumonkey 8y agoI'm still waiting for readings on n-SDV interactions too. I can sense the emergent resonance failure coming.
- YeGoblynQueenne 8y ago>> Waymo has “driven” more than 5 billion miles in its simulation, according to the company. Impressive. That's roughly equivalent to about 0 miles driven in real-world conditions.
- bob_theslob646 8y agoI'm assuming sarcasm. What other way would you have them test besides current ways that they are doing now?
- brudgers 8y agoTo me, the very existence of a self driving vehicle is an amazing technical achievement against a really hard challenge. I suppose that explains why the math makes sense. 25,000 aggregate miles/day for a fleet of 600 vehicles is below 42 miles/day/vehicle. Each vehicle is providing a level of service below 2 miles per hour. Again, as a geek I think it's amazing but not convinced it is going to scale to the point that will justify those governors pushing public policy changes that impact the general public...at least for now.
- lewis500 8y ago1042 miles per hour is a fast car
- oddmind 8y agoAre Waymo using AI and humans to build a ginormous curated decision tree that can be audited or are they doing something more black-box / pure AI / something else? I always imagined a big decision tree that could get diff’d after each AI training run and tuned by an army of humans.
- ebikelaw 8y agoWhy do you assume there's any "AI" in it at all?
- IshKebab 8y agoBecause this would fairly obviously be impossible without ML techniques.
- tim333 8y agoThere's some info on the systems here https://techcircle.vccircle.com/2018/05/09/how-google-s-waymo-is-using-ai-for-autonomous-driving https://techcircle.vccircle.com/2018/05/09/how-google-s-waym... >“While perception is the most mature area for deep learning, we also use deep nets for everything from prediction to planning to mapping and simulation.” ...Dmitri Dolgov, CTO and VP, engineering, Waymo etc