Having ridden in a Waymo in Phoenix a month ago, I don’t think they are anywhere close to being able to handle Manhattan traffic. As an example, during my ride the Waymo driver missed turns several times forcing a reroute around the block because it couldn’t change lanes because the driver in the next lane didn’t let it in. A human driver would have sped up or slowed down to get ahead of or behind the blocking car or they would slowly inch over the line to force themselves in. The Waymo car just maintained its speed and lane position until it was too late to turn. I can’t imagine how the car would fare in NYC with its famously aggressive drivers.
Sounds like a great reason to train the model on a closed road with a trained (and consenting) driver playing the role of "lane-blocking-driver" (in that example).
They also train in simulation, where they modify real scenarios to run through variations and edge cases. [1]
As of last year, they'd driven 15 billion miles in simulation - which is more than a typical person will drive in a million years. Still a ways to go to get to a billion though.
Or you could have a large fleet of
a million plus instrumented vehicles gathering data as they drive, then you could use that real world data to analyze and learn from all sorts of situations.
Sounds more like a great reason to stay away from some of the most chaotic traffic in the US until it can be considered safer. There are far, far more pedestrians in NYC than in Phoenix, and I do not trust a Waymo car to not hit some of them.
Well I think "high quality" is doing a lot there. Obviously autonomous vehicles would be safer if they work perfectly -- the concern is that they will not.
Yeah, approximately 10 minutes for me to pass driving test (from second attempt). And then around a year of driving (effectively training myself) in city to get more or less comfortable with traffic.
Generally, self driving cars are very safe, too safe actually. There are some high profile accidents, most of them involving Tesla because of the scale and way they approach the problem, plus an infamous one by Uber, but generally, self driving cars will not go anywhere unless they are sure the way is clear.
And this is the problem, self driving cars don't force their way, and they break when anything is vaguely in a collision course. If there is a problem putting them in NYC traffic is that they will stop all the time and create traffic jams, maybe get rear ended by other (human) drivers and fail to get their passengers to the destination because of some minor obstruction.
Worst case, add one more person armed with a crowbar and then we have the smash and grab thieves that seem to rampant in San Francisco with a very compliant automated driver stuck behind the accomplice just standing in front of the vehicle.
Gathering more data on a problem you don't know how to solve won't give you a solution to solving it. We don't have any evidence to support that current AI is capable of the advanced problem solving the city driving requires.
Every city on earth thinks their traffic is the most chaotic. Having driven around Manhattan a lot, I'd say it is probably the easiest use case for automation. Wide streets, perfect grid, slow traffic, flat, standard weather conditions, no two-way traffic lanes to cross. In fact I fail to see how this data would even be valuable to Waymo compared to what they are already getting from Phoenix and San Francisco.
You are either a troll, or have never driven a car in Manhattan. It is without a doubt the craziest, rudest, most agrrssive city I have ever driven in, and is 100% the hardest city to automate driving in. If Waymo can succeed in NYC, they can succeed anywhere in the first world. India and Thailand are a whole different ball game, but we will see.
I found driving in Manhattan to be fairly civilized, honestly. Boston, on the other hand, is by far the worst city I've driven in on the surface street level. Atlanta the worst for highway driving.
Yes yes, never mind the pedestrians for whom crosswalks and walk indicators are mere suggestions, or the thousands of delivery guys going 30mph on e-bikes in every possible (different from legal!) direction of travel.
And while there are lots of wide, slow, neatly gridded streets, there are tons of streets that are none of those things.
NYC is certainly not the most chaotic, but it’s a billion times more so than Phoenix.
It's worse than that. Not only they don't jaywalk, they stand or wander randomly all over the sidewalk/corner, getting in the way of those who are trying to get somewhere.
Downtown SJ might be the winner, though. Even at noon, in broad daylight, you might not see a car for a quarter or half mile (SF is never that deserted), yet pedestrians wait for the walk signal. Craziness. At least they don't give jaywalkers the stare of disapproval you'd get in Japan.
Only above 14th Street, and even then, that ignores e.g. Broadway which has a bunch of really frustrating-to-drive intersections with a ton of pedestrian traffic
> standard weather conditions
A year in Manhattan has significantly more variability in weather conditions than a year in Phoenix or San Francisco
> no two-way traffic lanes to cross
There are undoubtedly significantly more one-ways in Manhattan than many other cities but two-way streets absolutely still exist, most of the wider streets in the grid system are two-way, so's Houston
It isn't a perfect grid and the streets aren't all wide. There is no shortage of two way streets where left turns are impossible without skirting the law.
Agree, it's perplexing that Waymo thinks it makes more sense to start mapping out NYC rather than making the product work in Phoenix or SF.
You could interpret this as confidence, i.e. they're so close to a working service in Phoenix/SF that they want to lay the groundwork for NYC right away.
Or you could interpret is as a lack of confidence, i.e. they don't have much progress to report, and they're positioning that as "the reason we don't have a working product yet is that we're trying to solve the whole problem at once, and that's really hard."
> As an example, during my ride the Waymo driver missed turns several times forcing a reroute around the block because it couldn’t change lanes because the driver in the next lane didn’t let it in.
This is basically how I navigate NYC on my bike. Sometimes I can easily make a left turn, and do. Sometimes I evaluate that that's not possible, and make three rights at the next block instead. It works out in the end.
A human driver would have sped up or slowed down to get ahead of or behind the blocking car or they would slowly inch over the line to force themselves in
Is this exactly the kind of behavior that would make self-driving cars safer?
It will be safer but it will also be stopped in one place for a long time, possibly forever. A driver attempting to proceed through an intersection in Manhattan needs to present a credible appearance of being willing to kill someone, otherwise the pedestrians and other cars will just block it in a continuous parade, regardless of the state of the signals. I don't really see how you can program a car to deal with that.
> I don’t think they are anywhere close to being able to handle Manhattan traffic.
That's why they are starting to map new york and record human driving.
> Our vehicles will be manually operated by autonomous specialists at all times, to help us scale and advance our technology in support of our mission to make roads safer.
They are not starting any sort of taxi service, nor are they letting the computer drive, according to the article.
As a Massachusetts driver I think it would be great to get as many of these on the road because it would be free flowing traffic for me without having to deal with people being pissed off.
All the things I don't do because a human is driving I could start doing.
Drive all the way to the end of a line of cars taking an exit ramp and just cut it. The waymo car will always yield. Same goes for zipper traffic, just go for it, waymo will back off. Could probably also steal a parking spot from a waymo vehicle...
