This should be interesting. I'd like to see the "driver profiles" created for each city, assuming that's possible to create from observing traffic in different cities.
It's an interesting point. The level of aggression needed to get cross-town in Manhattan at rush hour is clearly very different from how you drive in a smaller city with modest traffic.
I'm still of the opinion that this is pretty far out except maybe on highways but good to see experimentation under a variety of conditions.
Very interesting—a sort of geographic and time-of-day based gradient across the parameters of the vehicle’s “profile” would be something very intriguing to hear more about.
Cruise Automation estimated it would take them about 6 months to learn a new city in preparation for a commercial deployment, which would include mapping, learning the quirks of specific intersections, signage, and presumably adapting their planning algorithms to local driving customs.
It seems like Waymo is well ahead of everyone else, but they had a mass talent exodus 2 years ago (many went to Nuro, Aurora, Otto, Argo, etcetera). I can't imagine that many smart engineers, already making good money even by Bay Area engineering standards, would quit to compete unless they really thought they had a chance.
There's just so much about the competition in this space I find fascinating.
Many were bored - they'd been working on it 5 years already.
Some thought the direction was wrong - not gathering enough test data (unlike Tesla).
Some thought they didn't have a big enough slice of the pie, and by moving to a startup and replicating the work they could earn way more through an acquisition by a carmaker.
Some saw the tech as fairly old. Remember much of Waymo's code was written by Urmson 10 years ago. Probably all C/C++. Not very nimble and quick to extend.
Some saw the fundamental approach wrong. Why not use neural nets for path planning and decision making rather than hand coded rules which will never cover enough cases to be good enough?
Some didn't want to be part of Googles risky approach. Too much pressure to get the cars on the road could kill people, and killing people, or the threat of it, is rather bad for morale. The leaders had this "even if it isn't perfect, it's better than current drivers" approach. Thats little comfort if it was your off by one error which caused the car to plough into a group of schoolchildren.
> Some saw the tech as fairly old. Remember much of Waymo's code was written by Urmson 10 years ago. Probably all C/C++. Not very nimble and quick to extend.
There was a rumor that Urmson stole code from the Urban Challenge and took it to Google, but Carnegie Mellon and GM never pursued a lawsuit. Or were you just being lax with your timeline?
Did you work at waymo or is this speculation? Some seems a bit like a Markov chain built from random self-driving objections ("never cover enough cases", "schoolchildren"), and some seems wrong (e.g. [1] mentions an expanding use of deep learning at waymo starting in 2013, well before Urmson left).
The usual big corp objections -- unsatisfied with agility, level of control, level of payoff and recognition if it succeeds -- seems a lot more likely.
> Some thought the direction was wrong - not gathering enough test data (unlike Tesla).
FWIW if you gather a ton of data and don't have simulation it's pretty much useless. Also it's expensive to keep around large amounts of data. Tesla didn't even have a simulation team until relatively recently.
> Some saw the tech as fairly old. Remember much of Waymo's code was written by Urmson 10 years ago. Probably all C/C++. Not very nimble and quick to extend.
Do you work in autonomous vehicles? Most teams are using C/C++. If you're not waymo you probably use ROS.
If you've ever worked at Google you'd know their C/C++ stack is among the most pleasant to deal with because much of the bad cruft from those languages is abstracted away in stable libraries.
Nah... Waymo internally had exponentially increasing bonus structure to achieving milestones [1]. This meant starting around 2012, lots of people got mind blowing bonus numbers. It was then soon become clear to top talent that they had enough money to either retire or self fund their own startups or join someone else’s without having to worry about meager salaries of startups. During same time brass decided to water down the bonus scheme and the party was over. Talent departed with their big money already in the bank. Much of the talent also knew that even bigger payouts was on the horizon with car companies buying up any ex-Google startups they can get their hands on.
> would quit to compete unless they really thought they had a chance
I disagree. Autonomous vehicles were/are a gold rush, and there are companies with lots of money willing to pay Waymo's engineers to... "replicate"... their success at Waymo. Many of them had an expectation to be acquired by a company with no software chops but a significant stake in self-driving, and that's proved to be a reasonable assumption over the past few years.
For the engineers, it's kind of like chefs at restaurants leaving to start new restaurants that are the same concept, but don't have the same volume or brand recognition.
Yes, you have to start from scratch at the new company, but you also stand to profit much more (and do interesting, early-stage work).
I am hoping that partnerships with ridesharing companies help with the data collection aspect. If Lyft/Uber could lease a few cars from all/some of these companies, which would probably pay them to be driven around, it might be a win-win for both the consumers and the companies.
Also, faster data collection at scale might reduce the time taken to "learn a city"
I recall taking an Uber once and hearing the driver mention that he wished he could remotely drive the car. Was this idea ever taken seriously? I'm sure latency and creating the actual interface would be big problems.
[/digression]
From a political perspective, what's necessary until the safety drivers are no longer legally required? I believe there are a couple places that don't require them now, actually.
> From a political perspective, what's necessary until the safety drivers are no longer legally required? I believe there are a couple places that don't require them now, actually.
Essentially, the legislature of the state in question passes a law defining what's required for truly autonomous operation. That can be pretty much anything they want. The really interesting stuff starts happening when there's enough of these things driving around that we start getting accidents. How will insurance companies or the sponsor companies deal with that? What will happen in any lawsuits resulting from any accidents? I guess we'll have to wait and see.
It's a bit sad, but it's probably for the best that there's big companies with deep pockets and big legal departments behind this.
your visual system is finely tuned to be able to judge things like "how far away is that child in the crosswalk?"
you possess an internal IMU that provides instant and accurate feedback on acceleration and rotation.
you lose out on both of those when you perceive the world through any kind of display, whether that be a screen or even a VR headset.
think about your field of view when you look out the windshield of a car. including peripheral vision that's a 150° field of view. compress that fov to a screen, even a really good one. maybe three of them set side by side. can you still make out a street sign with 6" tall text from 200 feet away?
If they can navigate which of the 800 Peachtree-named streets the car is on and which one it should go down next (minding all the crazy one way streets in downtown) without crashing into anything or anyone, they'll be doing better than most humans.
As an Atlanta resident, I can say that they have chosen some of the most challenging roads in the nation. The roads in Atlanta are a classic lava pattern. Many are literally the cow paths from town to town.
And we have some of the most aggressive drivers in the US.
31 comments
[ 2.1 ms ] story [ 68.1 ms ] threadI'm still of the opinion that this is pretty far out except maybe on highways but good to see experimentation under a variety of conditions.
There's just so much about the competition in this space I find fascinating.
Some thought the direction was wrong - not gathering enough test data (unlike Tesla).
Some thought they didn't have a big enough slice of the pie, and by moving to a startup and replicating the work they could earn way more through an acquisition by a carmaker.
Some saw the tech as fairly old. Remember much of Waymo's code was written by Urmson 10 years ago. Probably all C/C++. Not very nimble and quick to extend.
Some saw the fundamental approach wrong. Why not use neural nets for path planning and decision making rather than hand coded rules which will never cover enough cases to be good enough?
Some didn't want to be part of Googles risky approach. Too much pressure to get the cars on the road could kill people, and killing people, or the threat of it, is rather bad for morale. The leaders had this "even if it isn't perfect, it's better than current drivers" approach. Thats little comfort if it was your off by one error which caused the car to plough into a group of schoolchildren.
There was a rumor that Urmson stole code from the Urban Challenge and took it to Google, but Carnegie Mellon and GM never pursued a lawsuit. Or were you just being lax with your timeline?
It's just a rumor though (and a largely forgotten one at that). I just found it interesting that this person may have nonchalantly unearthed it.
The usual big corp objections -- unsatisfied with agility, level of control, level of payoff and recognition if it succeeds -- seems a lot more likely.
[1] https://nytimes.com/2018/01/04/technology/self-driving-cars-...
FWIW if you gather a ton of data and don't have simulation it's pretty much useless. Also it's expensive to keep around large amounts of data. Tesla didn't even have a simulation team until relatively recently.
> Some saw the tech as fairly old. Remember much of Waymo's code was written by Urmson 10 years ago. Probably all C/C++. Not very nimble and quick to extend.
Do you work in autonomous vehicles? Most teams are using C/C++. If you're not waymo you probably use ROS.
If you've ever worked at Google you'd know their C/C++ stack is among the most pleasant to deal with because much of the bad cruft from those languages is abstracted away in stable libraries.
https://youtu.be/7Yd9Ij0INX0?t=9m25s
is this true? ros works well for prototyping, but is it actually used in prod?
[1] https://www.theverge.com/2017/2/13/14599186/google-waymo-sel...
I disagree. Autonomous vehicles were/are a gold rush, and there are companies with lots of money willing to pay Waymo's engineers to... "replicate"... their success at Waymo. Many of them had an expectation to be acquired by a company with no software chops but a significant stake in self-driving, and that's proved to be a reasonable assumption over the past few years.
For the engineers, it's kind of like chefs at restaurants leaving to start new restaurants that are the same concept, but don't have the same volume or brand recognition.
Yes, you have to start from scratch at the new company, but you also stand to profit much more (and do interesting, early-stage work).
Also, faster data collection at scale might reduce the time taken to "learn a city"
Everytime I see Waymo news I hope it's about that Phoenix launch and then am disappointed when it isn't.
[/digression]
From a political perspective, what's necessary until the safety drivers are no longer legally required? I believe there are a couple places that don't require them now, actually.
Essentially, the legislature of the state in question passes a law defining what's required for truly autonomous operation. That can be pretty much anything they want. The really interesting stuff starts happening when there's enough of these things driving around that we start getting accidents. How will insurance companies or the sponsor companies deal with that? What will happen in any lawsuits resulting from any accidents? I guess we'll have to wait and see.
It's a bit sad, but it's probably for the best that there's big companies with deep pockets and big legal departments behind this.
you possess an internal IMU that provides instant and accurate feedback on acceleration and rotation.
you lose out on both of those when you perceive the world through any kind of display, whether that be a screen or even a VR headset.
think about your field of view when you look out the windshield of a car. including peripheral vision that's a 150° field of view. compress that fov to a screen, even a really good one. maybe three of them set side by side. can you still make out a street sign with 6" tall text from 200 feet away?
And we have some of the most aggressive drivers in the US.
edit: spelling.