Ford's problem may be Argo. Ford spent $1 billion for Argo, and paid for a big building with the Argo name on top. But Argo didn't get their California license to test autonomous vehicles until early this year.[1]
Self-driving has had too many "fake it til you make it" startups. Cruise started that way, and suckered GM into buying them for $1 billion. Uber bought Otto, and we now know how fake that technology was. Tesla hyped their basic lane-keeper and car detector into an "autopilot" then repeatedly plowed into clearly visible obstacles and killed people.
(Despite all the blithering about edge cases, they haven't been big problems in practice. The serious Uber and Tesla accidents were not edge cases. They were blatantly obvious obstacles: a semitrailer, a fire truck, a fixed barrier, and an isolated pedestrian on an open road.)
Waymo, meanwhile, keeps plugging away, driving around, getting their level of disconnects down each year, improving their sensors, and doing large scale tests. Once in a while they get rear-ended. If they don't get killed by Google/Alphabet's attention deficit disorder problem, as happened to Google's robotics efforts, they're going to get this into production.
It's not easy, and it's not impossible. It's just hard, like television or xerography. Those took decades from first demo until they worked well. This doesn't fit well with the startup make-money-fast model. It does fit with the big-company R&D lab model.
I agree Waymo is doing the best here, but Waymo is nowhere near a real self-driving car. They can't make unguarded left turns. Or drive in the dark, or the rain. Much less the snow, or in construction, or on poorly marked roads, or roads that aren't marked at all. Or city roads that aren't laser-mapped to the centimeter. Or roads with many pedestrians. Or cyclists.
And even when they do make it work, how is money going to be made by someone like Waymo? I don't see a route to the supposed massive profitability that would justify the huge investments being made.
> I agree Waymo is doing the best here, but Waymo is nowhere near a real self-driving car. They can't make unguarded left turns. Or drive in the dark, or the rain. Much less the snow, or in construction, or on poorly marked roads, or roads that aren't marked at all. Or city roads that aren't laser-mapped to the centimeter. Or roads with many pedestrians. Or cyclists.
I work for Google, opinions are my own.
I don't know the specifics myself but I trust what you're saying is true and agree those are real problems.
Nevertheless, one of the best things to do when you have a really hard problem is to simplify the problem. It's not as though we have to have fully self driving cars before they're released.
I think what would make a lot more sense is some middle ground where we have certain sections of the road where self driving cars will be able to work well and only allow them there.
That was the idea with Google's little bubble car. It was supposed to have a top speed around 25mph and cruise around retirement communities. That seemed like a feasible goal. But it cost too much to make as a product. The LIDAR units alone would have put it over $100K.[1]
Voyage.auto [2] claims to be deploying such cars now. Or rather, their web site contains announcements from late 2018 that they were doing so. Later information seems to be lacking. It's a basically good idea, but they claim an on-site staff of 5 for three cars, so they are nowhere near this making financial sense.
A real problem with self-driving cars is the false-alarm rate for emergency braking. If you're conservative about crash prevention, every once in a while the vehicle is going to detect something it doesn't identify as safe and will brake hard. That's why Uber turned off automatic braking in their cars - "to reduce potential for erratic behavior."[3] False-alarm braking, or even strong braking conservative by human standards, limits customer acceptance.
Or properly merge. Or recognize proper behavior in right-on-red situations. Or prevent their "safety drivers" from stopping in traffic to pick up friends for joyrides.
> (Despite all the blithering about edge cases, they haven't been big problems in practice. The serious Uber and Tesla accidents were not edge cases. They were blatantly obvious obstacles: a semitrailer, a fire truck, a fixed barrier, and an isolated pedestrian on an open road.)
> It's not easy, and it's not impossible. It's just hard, like television or xerography. Those took decades from first demo until they worked well. This doesn't fit well with the startup make-money-fast model. It does fit with the big-company R&D lab model.
Amen, and to my view, that's probably a good thing. As much as I think driverless technology will be a boon when it eventually arrives fully, the interim is damn scary. I would much rather have a decade or two of intervening half-solutions that actually provide well understood levels of safety. If a company was willing to shoot for something less than the moon, we might have been on track to have specialized carpool lanes that support driverless conveyance (e.g. buses or taxis) within major cities and possibly within 30, 50 or even 100 miles of them, depending on the metro areas and funds to adopt them. The problem is, who is willing to fund research into or consider implementing at a city scale technologies that required infrastructure investment when there's all these players acting like they'll have a perfect solution in a year or two?
I would much rather have a more modest but well understood solution in place in 3-5 5 years and a good beginning for a stable and trusted fully autonomous system in 10-15 a than the broken promises, and IMO incredibly optimistic PR we get now that serves no purpose other than to steal market share of money and public perception.
Who is willing to fund research into or consider implementing at a city scale technologies that required infrastructure investment when there's all these players acting like they'll have a perfect solution in a year or two?
Yes. Volvo demoed a system where they fired strongly magnetized nails into pavement to create markers they could detect through snow. Good idea, but doesn't seem to have been pursued. Volvo pointed out that it would be useful for snowplow guidance as well as self-driving vehicles. It wasn't intended to be the only reference, just a hint for use in bad weather.
edge cases... haven't been big problems in practice
Waymo doesn't acknowledge any "problem" that doesn't result in a collision. They don't seek input regarding such cases. Heck, most reputable trucking companies at least have a "his is my driving?" referral number clearly visible.
I've seen three cases where a driver had to take evasive action to dodge a Waymo fail in the past month alone.
My personal experience with Waymo autonomous driving has been pretty good. Never seen them do anything poorly and I see them a lot. They're known in Mesa/Chandler Arizona as decent enough drivers, but they can take a little long for a left turn and they also follow the speed limit rather than the flow of traffic which is the normal Arizona case.
Our industry failed here in setting expectations. We had people like George Hotz showing how easy it was, and Elon Musk promising "autopilot", when in reality he has fancy lane following. These people knew exactly where these systems break down, but glossed over that fact.
I'm optimistic on self driving technology, and I work at a company which builds a part of that tech stack. Here, in the bay area, there are a number of companies that got large amounts of funding trying to do self driving with cameras and computer vision alone. They'll fail. Camera based perception is not a solved problem, and it won't be for a long time, which leaves the traditional robotics sensing approach of Waymo and Cruise. These guys will have something workable first, but it won't be safe enough to be considered L4 or L5 anytime soon. Everyone else is miles behind Waymo and Cruise when it comes to generalized self driving, while others are quietly making good revenue on things like parking lot shuttles, mining equipment, and shipyard hauling.
I would love to have an autonomous mode in my car which requires no human intervention, so that I can sleep, or go out for a few too many drinks and have the car bring me home. Working directly on this stuff, I can see that I won't have such a car within ten years, maybe even twenty. I sure hope it's available in twenty, since i'll be too old to drive safely.
You're right, it's possible, but very hard. The barrier to cross from L3 to L4 is very high, and part of it is demonstrating to regulatory agencies that you have a safe system in all conditions, which will take a very long time.
"Here, in the bay area, there are a number of companies that got large amounts of funding trying to do self driving with cameras and computer vision alone. They'll fail."
Probably right.
..."which leaves the traditional robotics sensing approach of Waymo and Cruise. These guys will have something workable first, but it won't be safe enough to be considered L4 or L5 anytime soon."
L4, with slowdowns when the data is iffy, is probably achievable with available technology. If you're willing to accept occasional hard braking, and situations where the vehicle stops safely and asks the driver to take over. Customer acceptance problem with that.
Ford was working on autonomous stuff before Argo though, from what I understand. Argo was bought I think in hopes that they could significantly advance their tech with startup ingenuity. It's a bit disappointing because I think it's a good testing platform to develop these vehicles in Detroit given our variety of weather and road conditions that would be great for applying all of this in non-ideal conditions, but all I ever see in this city are May Mobility shuttles.
I don't think Waymo is really all that much ahead. Their testing is largely still in very ideal areas. Arizona has been attracting firms not just because they passed some permissive laws, but also because it's a place with lots of sunlight and relativity mild weather compared place like the Great Lakes or the Northeast. Same goes for California. Their progress however looks much better because they're really aggressively testing this stuff in ideal real-life conditions, but I think even they are a long way off from deploying those vehicles to places like Michigan year-round (a state that also has allowances for autonomous driving).
Waymo can be found on the roads in NW suburbs of Detroit (Novi). Admittedly I did not see them much over the winter. Point is simply that they seem to be testing in the more ‘real-world’ conditions found in the Midwest. I have no idea what the progress is or number of autonomous miles driven but they do have equipment in the field in Michigan.
Hopefully as other companies advance they’ll also move to Michigan as an ideal test bed and bring the crossroads of tech and auto to Michigan.
Waymo hasn't been in any serious accidents yet because they have human safety drivers and because they don't put their vehicles in situations where they can cause a serious accident even with a safety driver.
In that sense, expecting that Waymo is doing better than the rest and it's going to get self-driving cars into production is a bit like expecting that a person who learned how to walk on a line drawn on the floor is going to perform tight-rope walking because they're more careful than all those other idiots who actually tried to walk on a rope.
The thing to keep in mind is that self-driving is hard, much harder than tight-rope walking. It's so hard that it remains unsolved, currently, and no amount of careful application of non-solutions will result in a solution.
Each accident you cite happened under supervision of a human driver. Without knowledge about accidents prevented by supervisors it’s difficult to tell how much worse ai is than the average human.
Self-driving has had too many "fake it til you make it" startups. Cruise started that way, and suckered GM into buying them for $1 billion. Uber bought Otto, and we now know how fake that technology was. Tesla hyped their basic lane-keeper and car detector into an "autopilot" then repeatedly plowed into clearly visible obstacles and killed people.
(Despite all the blithering about edge cases, they haven't been big problems in practice. The serious Uber and Tesla accidents were not edge cases. They were blatantly obvious obstacles: a semitrailer, a fire truck, a fixed barrier, and an isolated pedestrian on an open road.)
Waymo, meanwhile, keeps plugging away, driving around, getting their level of disconnects down each year, improving their sensors, and doing large scale tests. Once in a while they get rear-ended. If they don't get killed by Google/Alphabet's attention deficit disorder problem, as happened to Google's robotics efforts, they're going to get this into production.
It's not easy, and it's not impossible. It's just hard, like television or xerography. Those took decades from first demo until they worked well. This doesn't fit well with the startup make-money-fast model. It does fit with the big-company R&D lab model.
[1] https://techcrunch.com/2019/01/29/argo-ai-acquires-permit-to...