Uber takes its self-driving cars off the road after one flipped over in Arizona
qz.com
qz.com
Uh oh.
The California DMV won't yet allow testing of autonomous vehicles heavier than 10,000 pounds on California public roads. Otto is thus testing in Arizona. Looks like that was a good decision by DMV.
I wrote this on HN yesterday:
An important question is whether their system is smart enough to take evasive action.
It's becoming clear that there are two ways to approach self-driving. The first stems from the DARPA Grand Challenge, which was about off-road driving. For that, the vehicles had to profile the terrain, plotting a path around obstacles, potholes, and cliff edges. The GPS route was just general guidance on where to go. That's the approach Google took, as can be seen from their SXSW videos. Google also identifies moving objects and tries to classify them. With all that capability, it's possible to take evasive action if some other road user is a threat. The control system has situational awareness and knows where there's clear space for escape.
The other approach is to start with lane following and automatic cruise control, and try to build them up into self-driving. This can be done entirely with vision systems. That's the Cruise Automation and Tesla approach. This puts the car on a track defined by lines on the pavement, with lane changes and intersections handled as a special case. There usually isn't a full terrain profile; that requires LIDAR. So there isn't enough info to plan an emergency maneuver for collision avoidance.
This distinction is not well understood, and it should be.
If true, that would seem to mean that Tesla's self-driving car program is dead on arrival. Clearly emergency collision avoidance is not optional for a self driving car with mass adoption. That's a pretty serious claim to make.
Detection of other vehicles, as shown on the dashboard display, is far below the level Google has demonstrated. Often, there's a nearby car that's not showing on the display. This probably reflects that Tesla only has vision and radar forward; side sensing is just short-range sonar, like a backup warning system.
What self-driving software Tesla is going to ship in the future is not publicly known, but we do know the sensor suite they're currently using, and it's far below Google or Volvo level for situational awareness.
[1] https://electrek.co/2017/03/18/tesla-autopilot-2-autosteer-t...
"Safety third" is a burning man-ism. It's kindof complex to even explain the context for this, but it definitely doesn't mean what it sounds like.
An example of this would be: my friends and I build fire effects. Some of the guys on my team are professional propulsion engineers at a company you have heard of, and they build rockets. So when we're building little flame throwers, and they say "safety third", they're saying it as joke, yeah, but their version of "safety third" is "This is only 20x overbuilt, not 50x".
To somebody not familiar with how that term is used, it definitely does sound bad, though.
This also does not inspire confidence.
> Recode also obtained documents showing that Uber’s self-driving cars currently need to be handled by their human safety drivers roughly once every 0.8 miles.
That's really far from autonomous.
EDIT: Here's the Recode article:
https://www.recode.net/2017/3/24/14737438/uber-self-driving-...
and things just get worse. The cars only make it an average of 2 miles between incorrect sudden movements, down from 4 in January. What a joke.
Google/Waymo's California DMV disengagement report for 2016 reports 0.20 disengagements per 1000 miles, or one disengagement per 5000 miles.[1] Google's cars drove 635,868 miles autonomously in 2016, so there's enough driving behind this for that number to be meaningful. This is 4x better than 2015, incidentally. CA DMV accident and disengagement reports are worth reading, since they're one of the few objective data sources available on automatic driving. So far, nobody has been willing to turn a self-driving car over to Top Gear or Road and Track.
Google/Waymo is thus three orders of magnitude better than Uber, and about two orders of magnitude better than anybody else who tests in California.
[1] https://www.dmv.ca.gov/portal/wcm/connect/946b3502-c959-4e3b...
No kidding! Yet somehow I'm not surprised, as it has always been pretty obvious that these jackasses didn't know what they were doing. This just adds to the many reasons why I am fully 100% behind getting them and other amateur-hour efforts (like geohot's comma.ai) off the road right now, and keeping them off until they can prove they are safe enough to be out on the public streets mixed with human drivers, cyclists, and pedestrians. I live and work in San Francisco and walk pretty much everywhere I go. I don't need to be watching out for some shitty "self driving" car that has to have a person take over every 0.8 miles whenever I'm trying to cross the street. Fuck that. If they keep this up, someone is going to get seriously hurt or killed.
Uber grew up on saying fuck you to government regulations. And it worked for them, because the only real losers were taxi cartels, and as much as I hate Uber, taxis were ten thousand times worse (in SF), which is more or less widely agreed on AFAICT. But it fails at dealing with the problem of autonomous vehicles, which is an existential threat to Uber. This is why Uber won't exist a decade from now.
Google grew up and succeeded through throwing lots of engineers at problems and building close, incestuous alliances with the government and academia, to avoid undue negative attention. This has plenty of downsides, but it's definitely the only plausible winning approach for getting autonomous cars actually on the road.
The driver is fully liable, in the same way that you are liable if you hit a car with cruise control enable.
Uber is level 2 self driving. That means that, legally, it is no different from the already on the road cruise control self driving feature. AKA, "The driver is legally liable at all times".
Basically I'm getting super annoyed at people claiming self driving cars are right around the corner, which I've been reading for a decade, but it's always "just 5 years away". I would love to counter arguments with, "If you are so convinced why not bet me $1000?"
I think a fair wager would be, "On March 26th 2022 at 2am, 830am, 5pm, and 11pm, a self driving call will legally take seibelj from [his address] in Boston to [address] in Cambridge. The car will require no human intervention and it will arrive at the destination regardless of time, weather, traffic, and construction."
Thats difficult. People who talk about "self driving cars 5 years away" aren't usually talking about mythical level 5 self driving that is actually 40 years away. They are talking about level 3, which basically means, "end to end travel, in 90% of situations".
Just like how 1/10 times you might not be able to flag down a taxi cab, today, 1/10 times you won't be able to get the self driving taxi to go EXACTLY where you need to go.
But it will work for most situations, most of time, and thats good enough. (IE, imagine if 90% of taxi drivers or truck drivers were made obsolete. That is industry destroying)
About "5 years away", I've been reading the same dream of all cars being self driving for so many years. I distinctly remember an after-work party in 2010 where a coworker with a phobia of driving going on about how self driving cars would soon be driving him around. Everything I hear today has been repeated for so long.
So 10% of the time you pay the normal taxi price that you pay now, and 90% of the time you get the super cheap self driving price.
That's still awesome!
So all these humans are just hanging around waiting for it to snow so that they have the pleasure of driving other people around? The market will react instantaneously?
Or are they doing other gigs in the meantime, like delivering people's lunches, laundry, groceries et al?
Maybe they will have other jobs. Most uber drivers work part time anyway, so that's not any different.
Maybe you have to pay people a premium to be an emergency driver.
Or maybe people will just drive less when it is snowing. Most of the time when I am using transportation, I don't 100% NEED to go on the trip. Maybe Ill just buy groceries tomorrow instead of today.
But even so, if 90% of the time, transportation prices are near 0 $, that is still crazy distruptive.
The market can easily solve any of the problems you are bringing up, just with behavioral/price chances.
And trucking isn't even affected by any of the issues you brought up. An x% chance of being a day late is easily worth 90% shipping cost decreases.
"Level 3: Within known, limited environments (such as freeways), the driver can safely turn their attention away from driving tasks, but must still be prepared to take control when needed."
L3 on the other hand may be a good and useful system to, say, automate freeway driving but it's basically useless as a system for replacing taxis.
Having driven from my apartment (at the time) in Boston to my job in Lexington going through Cambridge every day for almost two years... I will NOT take that bet and I'm pretty optimistic about self driving cars :P
But, they don't need to replace all driving overnight. They just need to replace a portion of driving to have a big impact.
Although, I often wonder if driving like muscles can become atrophied and if we end up with a situation like: "the car will drive me in fair whether but I need to drive myself when it snows"... if my body will forget how to drive by the time winter comes.
Personally, just around the corner means 10 years in the top 50% of weather. Operating in a blizzard may be very far away, and I'm fine with that. Would I take this bet? Probably not, but I'm passively optimistic.
http://www.theverge.com/2016/11/27/13752344/alvinn-self-driv...
Seems odd that they'd make this move unless it was to avoid more negative press for continuing the program.
Add to the fact that Tesla have the richest dataset they have a huge advantage in this area one that will reap rewards and crash less that the competition. I think Uber and Volvo will be far behind.
What are the challenges behind creating a licensing test similar to the ones we take as humans(could be virtual)? This would be before allowing the software on the road.
Such things could include push bikes, road crashes, collisions, crazy Deathproof style driving, etc. How do self driving cars react to other drivers using their horns or ambulances for that matter? How does a police car pull over a self driving car?
Interesting to see it's flipped itself somehow, too, I wonder what the computer decided to do that it maybe shouldn't have done? I suspect Uber will keep very quiet on this one, as there's all kinds of assumptions to be made.
The examiner marked it as a minor mistake; He told me that I could get rear end for that. I didn't argue with him as he was the examiner. But I'd prefer 10 rear ends over one t-bone.
When implementing a self-driving car, I'm not sure if we can explicitly write code that is against the DMV guidelines. For example, human seems to slow down at the junction even with the right of way and drive 5mph over the speed limit.
Self driving vehicles will have the same profile as the existing automotive industry: high capital costs and commodity margins. The existing automotive mega-corporations already have the capital investment and distribution networks like ships and rail cars and lots and auto-carrier semi-trailers. Tesla is building some of that and the jury is still out on whether Tesla will make a meaningful dent.
Even the idea of Uber rolling out a self-driving fleet means a massive infusion of capital that is at odds with its current cost structure. And the existing automotive mega-corps can step into the ondemand business with a cost advantage in regard to rolling out vehicle fleets.
On the surface autonomous vehicles appear to be a long-term capital draining exercise. Do they view the first mover advantage as that beneficial in this case that they'd risk all their future growth to be first to market for something that they could, conceivably, get a better ROI for being "second" to market with a more durable option?
But I think the real reason is that everyone building self-driving cars today - including the big car manufacturers - will have little interest in selling to Uber. They will be running their own Uber-style Car-as-a-Service.
https://www.youtube.com/watch?v=RY93kr8PaC4
But that trademark Volvo understeer doesn't like to powerslide.
People dont change their habits just because new tech is available. People are not going to start buying or using self-driving cars just because tech is available. There is huge psychological and social environment around car driving - e.g. think about all legal paperwork you have to go through to start driving, learning how to drive, buying insurance, all traffic regulations. All this cultural/human factors are based on specific definition of what car is and they all rely there is human subject operating this car. If self-driving cars are really going to succeed all this huge extremely social framework would have to change. I don't think this will happen. I think there are simply too many strong human emotions and purely financial interests around cars.