Trump administration re-evaluating self-driving car guidance
reuters.com
reuters.com
We will face a recession again. Trump will repeal Dodd Frank, banks will start making riskier loans and obscuring them in hidden investment products, and one day someone big enough will pull out, causing the whole house of cards to collapse. And, Republicans and Democrats will do another TARP because the alternative is World Depression II. That is, unless Trump seizes enough power to prevent TARP II.
Oh, I beg to differ. From a hard-core libertarian perspective, the state has no right to exist, let alone impose arbitrary regulations on business owners. People should be free to trade at will and guarantees on safety should be handled through a contract between the people who are trading. To impose regulations on trade would be an act of aggression.
During vehicle operation, everyone using the road would need to agree on an acceptable level of risk, probably managed by the road owner, not a state. Even if the owner of a particular road refuses to do business with you, you would still be able to buy whatever vehicles you like and operate them on your own property.
Libertarian paradises aren't always theoretical. See early bitcoin, and the commonly cited Somalia. On the opposite end, known authoritarian-tightass state China currently has the most efficient capitalist economy the planet has yet seen.
That's a completely bogus argument- you're conflating anarchism with libertarianism in order to make libertarians look bad.
If something is already safer than humans, then it should be deployed in mass, NOW. Screw the laws. Or get rid of them.
A million people die every year due to cars. The market can fix this if government gets out of the way.
Every year we wait is another million people dead, due to terrible human drivers.
Introduced prematurely, it could set the tech back due to public outcry. A few more autopilot deaths in the US in this political climate could really hurt the industry. Chao is looking for an excuse to pull permits.
But not in Arizona, or any other red state.
Get the federal government out of regulating it, and let the states decide.
Then, when red states save thousands of lives, the blue states will follow.
Haha or vice versa! Depends who proceeds in the right way.
I don't believe there's definitively an answer as to whether Tesla (arguably red and anti-regulation, moving forward without much caution) or Google (blue, more cautious) will find more success. I personally think Google will, but even better than that, it's great that we have such a system that permits that level of competition.
There's no excuse for a self-driving car to run a red light, but human drivers do it millions of times every day and in most cases when we see it we just shake our heads and say "what is that fool thinking?!" But a self-driving car does it once and it's headline news for months.
"Do you know why I pulled you over?"
"No" lie #1
"Did you know you ran that red light back there?"
"No, it was yellow when I went through" lie #2
"Here's the ticket"
person shows up to court to fight the ticket by claiming their innocence, lie #3
Again, no excusing Uber's behavior, just pointing to a double standard. We fully expect human drivers to run red lights, and we fully expect them to lie about it. There's no excuse for Uber's response and no excuse for their software running a red light, but there's similarly no excuse for humans doing it. Yet humans do it all the time and no one really cares.
Like if we're going to report headline news every time a Tesla catches on fire, we should have breaking news stop-the-presses coverage every time a gasoline or diesel powered car catches on fire, too. But we don't because it's much more frequent, to the point where we almost expect it.
The real question is not "do self-driving cars commit traffic violations?" but rather "do self-driving cars commit traffic violations at a lower rate than humans?". That's what matters. Not that they're perfect, but that they're better in a measurable way.
But we're not talking about the act of running a red light, we're talking about the act of lying about it.
>It's a lot easier to do a software update on a poorly-behaving self-driving car than it is to re-educate poorly-behaving human drivers.
What makes you think this?
>"do self-driving cars commit traffic violations at a lower rate than humans?".
Indeed. So the companies behind self-driving cars should prove it.
I disagree. This is a corporation speaking with one voice to attempt to continue to run cars which have just demonstrated an inability to detect a crucial traffic signal. Think of it this way, in one example I run a red light and lie about it. In the other I make arguments in the public sphere to allow the legally blind to drive because it will benefit me financially. That's not scale, that's fundamentally different.
Then Uber should disillusion me, not lie and cover up. Was it a flash of light off a reflection, a leaf that blew across the light, an odd angle? Then Uber should tell us that. More importantly though is the question of whether Uber even knows why it failed. What if they don't even understand why the car ran the light?
EDIT:
>You're assuming that because it missed the traffic light once it will do so every time.
Where did I assume that?
Where have I not been consistent?
>You're assuming that because it missed the traffic light once it will do so every time. The real world is a dynamic environment, something odd may have happened that it just missed the light that one instance.
Whether Uber misses this specific light over and over again or whether it fails in one off incidents across a spectrum of events when "something odd" happens shouldn't change our judgment of the car or company.
Plus, even if Uber's car recognized the traffic light as green, the fact that the cars before it on the other lanes stopped should be a clear enough signal to not just speed through.
For Uber, their car was probably not aware that the light was red, which is more dangerous.
I do not think "a self-driving car runs a red light" is analogous to "a person runs a red light". A closer analogy is to "a distinct class of people who run red lights".
I do agree that the media response is overblown, however I do not think it is without some merit.
https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...
Uber and all the other companies will permanently solve that problem the one time they figure out how to reliably identify a red light.
If a self driving car does the same people will think "ALL cars will make the same bad decission in the same situation". And there is some truth to that.
The human analogy is "<politician> ran a red light while drunk". The point is that they made errors of judgement that then caused this unsafe situation, rather than it being a random accident. (Especially in this case, in which high-handed defiance of both principles and the law is an established narrative.)
Humans fail in different ways, such as inattention.
that is really the hardest point, when you are trying to reduce unemployment like Trump is.
Never lower regulations when life are on the line.
There is a reason planes and cars don't crash every day, and it's because of these annoying regulations that force tests and QA. You'll never see a summer intern just come in and commits anything he wants and breaks everything. This shit doesn't happen in this line of business.
Regulations is good. You don't want devs being involved in your car with zero regulations to back them.
As a "tech guy" myself I have trouble fathoming a position of skepticism around the benefits of self-driving vehicles.
I know it is generally a good idea to be skeptical even as a "devil's advocate" but to my mind, there are so many upsides that it is hard to connect with me?
With that many in the sector its going to have huge economic consequences for the country.
See also: support for coal industry.
Take for example at least one prominent person who likely would still alive today if all cars were self driving https://en.m.wikipedia.org/wiki/John_Forbes_Nash_Jr.
Next would be how to measure the impact into real numbers like GDP.
It is a bit like terrorism, it is a lot unsafer to step in a wet shower in the morning or well, drive to work, than say getting killed by a terrorist yet look at the wars, waste, money, time spent talking about it.
The trolley problem would never happen in real life.
In all situations, the safe thing to do in an emergency is to slam on the breaks, and do avoidance, maybe. As well as don't get into these emergency situations in the first place.
Now we just have to convince everyone of that.
These states can be the test run of the technology, and the data they get will prove things to be safer.
Traveling at 50mph on an interstate in dense, almost no-visibility fog. Slamming on the breaks is going to create a 50-car pileup with at least two or three fatalities. You might say it's not safe to be driving that fast in those conditions, and you're probably right, but everyone else is, and driving significantly below the going rate is equally dangerous.
There is a huge psychological element to why cars are so popular.
[Personally I hate cars altogether and only use public transport.]
Also:
Self-driving cars won't start out safer than humans.
Cars which can self drive in some situations but not others will be abused by lazy humans.
Self-driving cars will drive differently than people and thus kill people in accidents which would not have occurred had there been a human driver.
There is no such thing as "mostly autonomous." Humans will not be able to take over and drive at their normal skill level within short time frames.
There has to yet, been no apples to apples comparison of human drivers and autonomous vehicles. Until we get that, why should we let companies use our public roads as their testbed? When Tesla or any other company can prove to me that their cars are safer than human-driven vehicles, I won't oppose them on my roads.
All data that I've seen so far supports the conclusion that self-driving cars are already much, much safer than humans.
http://bigthink.com/ideafeed/googles-self-driving-car-is- ridiculously-safe
> When Tesla or any other company can prove to me that their cars are safer than human-driven vehicles, I won't oppose them on my roads.
How can they be expected to prove their safety record to you without driving hundreds of millions of miles on public roads? They're already provably safer than humans in artificial "test course" scenarios.
>How can they be expected to prove their safety record to you without driving hundreds of millions of miles on public roads? They're already provably safer than humans in artificial "test course" scenarios.
If they are so ingenious as to build a self-driving car, you are telling me they can't think of a way to actually do an apples to apples test? Or is it that they don't want to prove their cars are safer? Or is it that they don't want to pay to prove their cars are safer?
Tesla is already less-safe when looking at fatalities involving autopilot. There have been one (FL), probably two deaths (China). That was when they only had around 150 million miles driving mostly in good weather and on divided roads with clear markings.
The 1 in 100 million average in the US includes motorcycles and all types of cars, new and old. Tesla is a new luxury sedan driving on limited roads in limited scenarios.
I don't have trouble being skeptical, in fact. I don't believe it will ever work for city streets.
Here's why:
The computer would have to be 99.99999% reliable to do that, and we don't have computers that reliable.
The numbers:
The accident rate is around 74 per 100 million miles (and fatalities is 1.13).
It's unclear exactly how to turn that into a percentage, but no matter how you do it it's quite high.
Say an accident takes 5 minutes, and people drive 30 miles/hour. Then that works out to 99.999% for humans. If you use the numbers for fatalities then it's 99.99999%.
I.e. 99.99999% of the time, as whole across all [US] humans, people drive in a way that does not cause a fatality.
That's the bar computers have to cross in order to save any lives at all.
Here's a less bitter contrarian view:
Self-driving vehicles will be bad for American city development and reinforce existing patterns of inefficiency, waste, and self-segregation.
The reasons I expect this to happen: * People will likely put up with longer commutes if they don't have to pay attention the whole time. * The amount of cars on the road for a given level of congestion will probably increase due to more efficient cross-vehicle coordination and faster reaction times. * I expect most consumers will still prefer to own rather than perpetually rent their vehicle (the average age of cars on the road right now is 11.5 years - that's quite a few years of not making payments compared to being a perpetual Uber-rider), especially given the convenience in family situations (keeping stuff in the car all the time to entertain the kids, keeping the back seats covered to protect from the dog fur and claws, etc).
So those plausible resulting conditions seem likely to lead to less centralized cities, reinforcing a future of inefficiently using a bunch of energy to move small numbers of people individually in cars vs in a denser, older-style urban rail system.
Cars gave nothing to do with it. The problem is that 10s of millions of people would live in the city if they could, but the government does let them, by putting so many limits on house building.
Cities don't have rights. Why are you equating "good for cities" with absolute good?
"and reinforce existing patterns of inefficiency, waste, and self-segregation."
What you call "inefficiency, waste, and self-segregation" other people call "living the lifestyle they want, having a nice yard, and being able to raise their kids away from the dirt, crime, and horrible schools found in the cities".
Efficiency is not the end goal of human life.
If we were really after maximum efficiency, we could have everyone sleep in bunks (maybe even hot-bunk it with three shifts) and eat all their meals in communal chow halls.
But we don't. Why is that, do you suppose?
Considering traditional automakers previous behavior when dealing with defects, this huge rush to get cars connected so that they can hoover up as much data on you and your driving as possible, their general culture of secrecy, a reticence to launch bug bounties and the general closed source nature of their code, you are not at all worried about terrorists or, much more likely, mischief makers hacking your car (or a fleet of the same model with the same vulnerability) for the lulz?
I've said this before, but I'm definitely not getting Mk. 1 of a self-driving car and most likely will never get one unless they open source the code.
California DMV's regulations for testing are straightforward. 1) Get $5M in insurance, 2) Report all crashes immediately, and disconnects annually, and 3) only trained drivers associated with the manufacturer can drive. That's what Google wanted, and there are about 20 manufacturers with CA autonomous vehicle testing licenses. Their reports are one of the few objective data sources we have on self-driving cars. (Uber whined about having to report their failures, refused to register, and moved testing out of CA.)
California's deployment regulations are still in flux, although there's a draft version. So far, nobody has applied to deploy production self-driving vehicles in California.
NHTSA started with definitions of levels of autonomy. Informally:
- Level 0 - manual driving
- Level 1 - auto brake or lane keeping, but not both.
- Level 2 - auto speed/brake control and lane keeping (Tesla, BMW, Mercedes, etc.) Driver may have to take over control at any time.
- Level 3 - fully automatic driving on some roads. Driver may have to take control, but not instantly. (Google/Waymo, Volvo)
- Level 4 - fully automatic driving on all roads.
- Level 5 - no driver needed.
NHTSA proposes that at level 3 and above, the manufacturer is responsible for all accidents. Volvo is on board with this, and the US Big Three automakers don't object. NHTSA sees a bright line between levels 2 and 3, and says that at level 2, manufacturers should enforce hands-on-wheel. That's what got Tesla into trouble - weak hands-on-wheel enforcement combined with PR that convinced some drivers that "autopilot" was better than it is.
Urmson, when at Google, said that their testing taught them that humans cannot take over fast enough to deal with automation failures. This echoes what the aviation community has discovered over the years - when someone who isn't actively flying has to take over control, it takes seconds or tens of seconds before they have full situational awareness.
This means the incremental approach of adding more features to assist the driver won't work. Somebody has to be in charge, either the human or the computer. Ambiguity over who is driving leads to accidents. (For an aviation view of this, see "Children of the Magenta", where a chief pilot is talking to his pilots about cockpit aviation dependency.[1])
Congressional hearing last week on self-driving cars.[2]
[1] https://www.youtube.com/watch?v=pN41LvuSz10 [2] https://www.youtube.com/watch?v=vJsRF7NcjpM
I'm disappointed that this didn't become the prevailing wisdom for self-driving cars. With the current state of affairs, it seems there's a real risk of a few accidents causing this administration to pull the plug on self-driving altogether until we get a progressive administration again.
Perhaps Urmson ought to have published this research so that others could use their methods and reproduce the results. It seems to me that user-study was the most important part of their research.
1: https://www.youtube.com/watch?v=uYav3_7miIc
2: https://www.youtube.com/attribution_link?a=zYHrJvXcz6M&u=%2F...
Tesla broke ties with the previous supplier of its driver assist hardware and software last year after serious disagreements over the safety of Tesla's present and future Autopilot systems and how it was being marketed to consumers.
It's now been just a few months, and Tesla's rolled out their own Autopilot software to customers of cars built since the breakup with Mobileye.
That software doesn't have the years of testing the old software did. This one, as you can see in the videos I linked from the latest firmware, does things like drive your car off clearly marked roads inexplicably without warning. It is not acting safer than human drivers would in these situations.
I'm not happy about sharing the roads with cars that may drive themselves into me with no warning that there's anything wrong with what should be a simple lane keeping system. Especially when that system doesn't force the driver to keep their hands on the wheel, when obviously they need to be there with the software in this state.
I want to eventually benefit from the kind of self-driving cars Google is building (Waymo is down to a human temporarily taking the wheel once per 5000 miles of driving on public roads), without the harm of what startups like Tesla can do by putting essentially untested systems on production cars I share the road with.
There must be a regulatory middle ground?
I think there is, I just wouldn't hold my breath for Congress to act on it. California is leading the way on this. Elect the right people there and hopefully that trickles up and across to other states.
Plus, the states are the ones regulating on this. California is leading the way in independent testing, and they and a handful of others are the only ones who've begun requiring reporting about accidents involving driver-assist and autonomous vehicles.
Am I dumb for saying this administration is unlikely to make federal laws surrounding this technology? Their whole argument is to shrink national government and empower the states to do as they please. Every time they make a national law they shoot themselves in the foot.
Agree, I believe that's a fair assessment and I am glad they are recognizing it.
Flagging is totally sick on HN.