California DMV: Self-driving cars must have driver behind wheel | Hacker News Reader

California DMV: Self-driving cars must have driver behind wheel

mercurynews.com

42 pointsnot_that_noob77 comments

Its nonsensical to think a human can take over at a moment's notice during a particular maneuver. Its the threat of death/accident that makes me pay attention to every moment of driving. On autopilot I wont have that kind of attention. Its simply a trait of the human attention span.

That being said, every self-driving car needs a manual override because there are plenty of situations that cant be easily explained to a computer. the override could be done with a steering wheel and gas pedal, or just a joy-stick mode on the touchscreen pad.

For instance someone has a big empty lot and they want to move a car from row A to row B. The new spot is exactly 50 feet to the right. I just want to back the car up, and go forward while steering over to the right a bit. Row B has no 'street address' to program into a route-computer. I could click on a map but the map doesnt know where on the grass i want to drive.

Even with lidar, etc. there are grey areas that self-driving algorithms are going to have trouble with. Drop a sheet of tissue in the path of a self-driving car moving at 60MPH. The lidar will see it as a solid object and might decide a very dangerous plan - maximum braking - is the best plan, when going through it is actually the best plan.

This already happens today with cruise control - and yes, the driver will have to be paying attention while driving, at least in this version of the rules.

Cruise control still requires one to mostly drive, and the only "moment's notice" thing a human would do with cruise control would be stomp the brake.

Regardless of what any rule says, stand-by humans will not be paying attention and ready to drive. Even if reading/sleeping/daydreaming were successfully policed, the human will still be out of touch with how the car responds to controls.

Think of the first time you drive after not having driven for a few months - now imagine being already moving at highway speed.

All cruise control does is set the speed, it is not analogous to self-driving.

You assume it's for safety, but it's more likely to be a case that the car can't handle at all and simply stops.

For instance, if there's some kind of minor obstruction, a human may decide to drive over it or go into the oncoming lane or get out and pick it up or follow an officer's instructions.

CA DMV's position is basically "let's try it with human backup for 3 years, then re-evaluate". That's reasonable enough.

Tesla may have set the field back. They released their "autopilot" as a beta, didn't provide sensors to enforce driver hands-on-wheel, shipped a system that works well only on freeways but will engage on other roads, and claimed that if it crashes, it's the driver's fault. Everybody else shipping similar capabilities (BMW, Mercedes, Cadillac, etc.) has put more restrictions on them (such as a hands-on-wheel sensor) to avoid over-reliance on the automation. Now Tesla has had to remove some features from their "autopilot".[1]

Google may be able to get approval for their 25MPH max speed mini-car operating without driver hands-on-wheel, on the basis of the slow speed. Google's system is much more advanced, and has a detailed model of what everything else moving around it is doing.

Expecting the human driver to take over if the automation fails will not work. The limits of human monitoring are well known from aircraft automation. It takes seconds to tens of seconds for a pilot to grasp the situation and recover properly if the autopilot disconnects in an upset situation. That's for pilots, who are selected and trained much better than drivers, and who go through elaborate simulator training in which failures are simulated. Watch "Children of the Magenta"[2], where a chief pilot at American Airlines talks to his pilots about this.

[1] http://www.bizjournals.com/sanjose/news/2015/12/16/tesla-to-... [2] https://www.youtube.com/watch?v=pN41LvuSz10

Also pilots generally have much more time to respond to a automation failure. If the automation fails in a car, even a 1 second driver response could not be enough.

Planes are more complex than cars. Drivers react to road surprises far faster than 10s, when making a decision to brake or swerve

That reaction is not necessarily right, its just reflex which can produce more disastrous results than no reaction at all. Can't trust them much, particularly when I'm not fully aware of the situation around me (and I think one is unlikely to be fully aware of situation when on self-driven cars).

> That cautious approach requires that the cars have a steering wheel, and a licensed driver must be ready to take over if the machine fails.

The article takes the tone that this will drastically limit the technology, but honestly at this point a restriction like this seems completely reasonable.

The cars are not good enough to completely drive themselves in all scenarios. At times, they will decide they can't handle a situation and safely ask a driver to take over. This is becoming more and more rare, but it's definitely still the case. Until the technology is proven to basically never need human intervention, it seems reasonable to require a person who can drive to be present.

For cases such as if the car detects a malfunction, this probably won't be a big deal, but for most extreme situations where a human will need to take over, and that can't be fixed by stopping the car, will a human be able to take over in a timely enough manner to solve the issue? We are often talking split second decisions, and it will involve a human who likely wasn't paying any more attention to the road than a passenger.

Imagine a car where the passenger's side is fully set up to drive, and with the push of a single button, the driver can swap all control to the passenger. I'm having a hard time thinking of situations where this is better than just stopping the car.

Hmm, that's makes me wonder how driving lessons work in US. In Poland it is a legal requirement that the car used for such purpose has a second set of pedals on passenger side, where instructor sits.

When I took a driving lessons my instructor was perfectly capable of pressing the brake or overtaking the wheel on time when I was about to do something stupid. I don't see why manual override of autonomous car should be harder.

I see however, that this can easily defeat purpose of using autonomous car: if the driver must be fully concentrated all the time, be ready to react, then it is even worse than driving normal car.

It works the same way in the US. However, in that case, the instructor is normally paying attention to the road and the student's driving, to grade them if nothing else. Usually, passengers can't be relied on for that level of attention.

When I took driving lessons in the US, the instructor had a brake on the passenger side but no other controls.

That's a great question, and my understanding is that the car simply doesn't ask humans to take over in that case.

The "handoff" technique is primarily useful in situations where the car's confidence in the situation slowly degrades, not for sudden collision avoidance. An example would be a confusing construction area, or navigating with rocks on the road after a landslide. In these areas where road lines and signs aren't as applicable, the car needs a human driver. The human driver is so people aren't abandoning cars at the start of construction zones, not so they can step in to avoid a sudden collision. Google recently did a study and determined it takes over 5 seconds for a human to take over in a scenario, the car won't be able to expect them to handle anything that is happening quicker than that.

One such situation: Car detects malfunction while in an intersection. If it comes to a sudden stop, then at the least it will block the intersection, and you'll have the passengers sitting in a box in the middle of the intersection struggling to get the car going.

Further, a malfunction while in the left lane of the highway would be just as dangerous.

I don't even drive that much, all my commuting is done by bike. These are just a couple things I can think of off the top of my head.

> Imagine a car where the passenger's side is fully set up to drive, and with the push of a single button, the driver can swap all control to the passenger. I'm having a hard time thinking of situations where this is better than just stopping the car.

I don't know about you, but when I'm a passenger in a car, I'm attentive to everything. I guess the fact I often have strong rapport with the driver (usually a family member) means we are in sync; I mean I'll literally do blind spot checks, and carry out many of the driver's visual rituals simultaneously.

I think this is an interesting scenario; machines assisting men in, for example, collision avoidance systems, or in this case all the way up to self driving; and men being required to help self-driving machines in unusual situations.

Drver heart attack aka computer sensor failure.

> The article takes the tone that this will drastically limit the technology, but honestly at this point a restriction like this seems completely reasonable.

This does seem like what the first step to getting self driving cars on the road would look like.

For a time, why not allow the cars in fully-autonomous mode on a restricted set of roads, even particular lanes on particular roads, at particular max speeds, marked as such so everyone, including the cars, knows what to expect? In the case of unusual construction particular routes could be temporarily disabled. This would severely curtail the number of outlier cases the software has to be ready for, be actually safe, and allow for awesome quality of life improvement for commuters on those routes who can do something else while commuting. And in the meantime, gain huge number of hours driving experience for the software, enabling the fully autonomous routes to be slowly expanded over time.

The large part of the complexity of implementing self-driving cars comes from the fact that the roadway is not a fully protected space. You can find a number of examples of fully unattended automated vehicles that have been around for a while (for example, AGT people-movers that are common at airports), but they are dependent on a fully protected pathway so that they do not require very much intelligence.

So what about roadways? Freeways have a certain degree of protection, but it's still not comparable with an elevated train ROW. It is certainly possible to develop roadways with a sufficient degree of protection for less sophisticated automation to be reliable, but it would be very expensive: they would need much more intensive maintenance, traffic density would have to be quite a bit more limited, they would need careful physical protection with at least fencing. In the end, I don't think that the infrastructure investment would be worth it considering that the period of time for which this infrastructure is needed will probably be quite short (in infrastructure terms).

How will this impact one of the key arguments for self-driving cars: elimination of DUIs?

If a driver is required behind the wheel able to take over at a moment's notice then the self-driving car's purpose is defeated - you cannot do work while the car drives, you can't use the car as a sober ride home from the bar, you can't take a nap on your commute, etc.

This is a legislated death sentence for self-driving cars as we envision them.

While I agree that it defeats many of the purposes (including Uber-style services), remember that this legislation is in response to current state of self-driving cars. Once they get better, legislation can become more lax.

As a general rule of thumb, legislation never becomes more lax.

That's a silly thing to say. Uber and AirBnb were illegal when they launched. You couldn't use your phone on an airplane during takeoff until recently. The original speed limit for cars was "no faster than walking".

With the lobbying power of Google, Tesla, Apple and all the car manufactures? It will change.

Actually it most likely will lag years or more likely decades behind the capability of the cars because the people that stand to profit from such rules limiting the usefulness of self driving cars.

Little soon to be doomsaying.

First generation self driving vehicles might have these types of limitations, but second generation? Third generation? Ten generation? It is all about building trust. Truth for the general public, and those who will be held responsible if something goes wrong.

Aircraft still, to this day, have this exact system. Ask any pilot if the system is pointless simply because they have to be paying attention? It still requires less mental energy than actually driving, in particular for longer road trips where the mind wonders.

The main issue with the legislation is that it's slow to overturn it once we're ready.

Which is fine. Trust is built slowly and the public should have a chance to debate it.

If most of the public thinks it is "common sense" (by that stage) that self-driving cars are safe, then we're in a good place to be overturning it.

This is one place, public safety, where "move fast and break things" doesn't work because those things will be literally people.

We're talking about driving here; we already move fast and break things, that's the problem. Google's koala cars move slow and have a foam front end.

Even though it doesn't look like it, this is actually great news since this provides a path for Google et al to get out there and build broad customer comfort around the technology.

And frankly I don't ever foresee driverless cars that don't have some form of manual ovveride - there will always be situations that won't make sense to the computer that make perfect sense to me.

==== Edited to add this clarification, as people on HN seem overly ready to pick nits. ====

The larger point is that humans will be humans and will never buy a car that will interfere with their freedom to do what they think is right at that moment.

And if you want a better example - maybe I want to speed to get to the hospital in an emergency. Maybe I want to block another car that's leaving the scene of a crime. You get the idea.

> park in this no parking zone blocking traffic for 30 seconds while I run to get something.

I think you just made an argument against manual override there...

...until you learn that the reason he had to stop in a no-parking zone for 30 seconds is because it was the shortest, safest route for a severely disabled person to get into the car.

> park in this no parking zone blocking traffic for 30 seconds while I run to get something.

No? Either park legally or go without. You are literally saying to everyone else "me saving 5 minutes by not having to park is more important than whatever it is you're doing."

You realize it's a hypothetical right?

The larger point is that humans will be humans and will never buy a car that will interfere with their freedom to do what they think is right at that moment.

And if you want a better example - maybe I want to speed to get to the hospital in an emergency. Maybe I want to block another car that's leaving the scene of a crime. You get the idea.

Sometimes rules need to be broken. I don't want to live in a society where it's impossible to break any rules, and neither do you.

Those who disagree would probably be happier on some other site.

Does your microwave let you override its safety choices?

Sure, if I stick a screwdriver into the door interlock, or peel off the choke ring seal. Why?

Actually since the "driver" had not left the car, it would be considered "standing" not "parking", at least by MA law.

Only if the "driver" can decide to move out of the way if needed, without the passenger directing it

> park in this no parking zone blocking traffic for 30 seconds while I run to get something.

You're not thinking in driverless terms. You don't need to have the car block traffic while you grab something. The car can drive around the block and swing back around to pick you up or just go find a parking be spot elsewhere while you do what you need.

Um, describing why you want to be an asshole to everyone else on the road isn't a good way to get people to agree with you.

Actually, Most people relate to this behavior and rationalize it as just 30 seconds!

Citation? I suspect most people are somewhat conscientious drivers who operate their vehicles predictably most of the time. But this is just anecdotal from my own driving experience.

Are you a sentient neural network? If so, ask any of the humans that you can interact with. That's all the citation you will need.

There is definitely a cultural aspect to this. Where I used to live, it would be insane to leave your car in a traffic lane and if the cops saw it they would certainly have words for you, if they hadn't asked for a tow.

In San Francisco, this is the normal way of doing business for many people, and while the left lane sucks because of people waiting for yielding left turns, the right lane sucks just as much because of all the cabs, uber/lyfts, delivery trucks, and random people doing god knows what who have left their car in the lane of traffic.

I figure that anywhere where parking is very difficult to find, it's going to be more socially acceptable to double-park, even when doing so blocks traffic.

What about situations where the person buying the car is not the person driving the car, like a company car or a rental agency? I could see it being cheaper (for insurance purposes) to not allow a manual override.

Sounds like a good idea for the first 10 or so years of self-driving cars. Remember, not all self-driving cars will be made by Google or Tesla. Some will even be made by the spaghetti-code-Toyota, the cheating-Volkswagen, or the we-send-updates-over-unencrypted-HTTP-BMW. It's best not to jump head-first (literally?) with this.

And it's not just about the "safety" code of these self-driving cars, either. Many of them probably still won't care too much about hiring top-notch security experts in the first few years, to protect their cars against remote hacking (which could be much easier with self-driving cars if they are connected to the Internet).

If I have to be responsible for AI and mechanical errors, I don't want an AI driving my car.

Either I want to read a book or take a nap, completely zoned out, or I want to be alert and paying attention to the road.

Sitting there doing nothing bored and disengaged, not able to read a book, not actively engaged in the activity of driving, and also held responsible by the state for failures is the worst of both worlds.

If that AI has, say, one-third the risk of accidents as I would have, then I want it driving my car even if I enjoy it slightly less.

It's not the I would enjoy the experience less, it's that the experience of just sitting there would render me an inattentive dullard with very poor response to a mechanical or AI failure. If I'm suppose to be responsible per California State, I need to be paying attention. Just sitting there, inactive, mile after mile, hour after hour doing nothing, maybe for months or even years at a time, I will be woefully unprepared to respond to the emergency.

I actually gave up driving for 16 years then restarted when I last moved. What I found when I restarted was that, while physically I could operate the car, my decision making skills when faced with a problem were non-existent. I could see problems coming my way, but the reflexive decision making was gone. I had to relearn the decision making process.

Designing the system to place responsibility on humans to take over only occasionally will ensure that the system creates catastrophic results when humans are found to be unprepared for the responsibility.

Brilliant idea on the DMV's behalf. I'm reminded of Douglas Adam's thought, in Hitchhiker's Guide to the Galaxy, that solving a problem ("Now that it's perfectly air-conditioned,") shouldn't allow the removal of redundancy ("...you won't need to open the windows, ever."). Because the problem won't REALLY be solved in all cases, and if the driver finds he's in an ugly situation where he cannot act, there's no steering wheel to handle the convenient killing machine he's riding. Can you imagine the horror?

> Can you imagine the horror?

About the same as something falls from the sky and crushes a child. Unlikely enough that I don't waste my time thinking about it. The self-driving cars are more cautious than in all situations and ceases to be a killing machine and acts like what it is. A transport device. Even with the possibility of error, I am supremely confident that a self-driving car is more competent than the average driver. The legislation is out of touch with reality, as usual.

We can include an emergency bike in the trunk so you can still get where you are going.

Bikes have the highest practical value of 100% computer-free objects, so this looks like a reasonable solution.

This would be a huge setback if approved simply because once approved, such things are unlikely to change even when the time comes that human drivers will not only be unnecessary, but the most dangerous component in all situations. I understand the state's motives: liability, fear of bad press, keeping traffic cops employed (and thus the biggest avenue for random arrests), etc. but none of them justify limiting the technology simply to serve their rather short-sighted, selfish needs. At the same time, the people that could benefit most from driverless cars like the elderly, disabled, youth, etc. are shit out of luck. And all this without a single accident yet attributed to driverless cars (afaik).

This is the exact wrong idea. Frankly, consumer-ready self-driving cars should be prohibited from changing into "manual mode" without a user request. And the manufacturer should be liable for an at-fault accident in auto-drive mode. Until these conditions can be satisfied, self-driving cars are simply not ready for prime time.

The proposed rule will lead to proliferation of shoddy "99%" software designed around an expectation that a human can take over at a moment's notice. But humans don't work that way, so any such exceptional condition will likely result in a crash. The only acceptable failure procedure for a self-driving car is to stop.

See my comment above on two situations where coming to a stop is actually more dangerous.

TL;DR: If the car fails while driving on a highway or crossing a busy intersection, stopping will cause a wreck.

Sure, coming to a stop is dangerous in those situations if other drivers don't react appropriately. It's not a great idea, but I also don't mean that self-driving cars should be doing such emergency stops regularly. And obviously they should do it only as an extreme last resort, when things like pulling over have failed.

I still think it's more reasonable to stop either of those two places than to suddenly hand control of a car traveling 75mph to a human who hasn't driven in a year or longer.

edit: concrete example: You're biking along. Would you rather A. a car in front of you come to an abrupt stop, or B. a car passing you at twice your speed be controlled by a drowsy driver who doesn't know where the edges of his car are and doesn't have a feel for how the steering wheel reacts ?

A fair point. The argument really comes down to this: Is it likely that a malfunction will occur while the auto-car is closely followed by another car?

Well, I wouldn't say "likely" comes into it. I'm explicitly describing what happens with a choke_and_die() function if say the software crashes with no recovery, or major sensors are damaged - unpredictable by definition. The "I don't know how to navigate this situation" case would be handled by gracefully pulling over, and should rarely happen because developers would be encouraged to work out the kinks rather than having such a terrible user experience be common.

My point is that giving developers an easy out of forcing a human to take over just encourages sloppy non-robust software. And if there's anything that consumer technology companies are good at creating, it's non-robust software that has no place in a car.

Does the requirement need to have a control unit at all times? It would be neat to continue to make it an option but also allow the console to hide the wheel and perhaps replace it with some more fly-by-wire smaller replacement.

Maybe not hands on the wheel control but I would at least want it to be a wheel that is ready to go by just grabbing it, same with the brakes. If they collapsed or folded into the dash that could lead to issues where a software glitch doesn't let you take back control and the robot overlords end you.

I feel like it could be more unsafe this way. What if a human gets scared, takes control, and overreacts when the car is properly handling the situation? Might cause more accidents.

EDIT: reworded sentence so it actually made sense :P

The human who drives a self driving car now should know what he's doing. These cars are prototypes now and whoever takes them to the road should be responsible.

When I read the geohot article [1] earlier I was thinking that there was no way what he was working on was legal. Would be curious to have someone with more expertise weigh in though.

Does this make comma.ai [1] or cruise [2] more acceptable in the eyes of the state?

1 - https://news.ycombinator.com/item?id=10744206 2 - http://www.getcruise.com/

California rules (summarized):

A manufacturer may conduct testing of autonomous vehicles on public roads in California if:

(a) The manufacturer is conducting the testing.

(b) The vehicle is operated by an autonomous vehicle test driver who is an employee ...

(d) Manufacturer’s Testing Permit from DMV.

Current California permit holders:

    Volkswagen Group of America
    Mercedes Benz
    Google
    Delphi Automotive
    Tesla Motors
    Bosch
    Nissan
    Cruise Automation
    BMW
    Honda
    Ford

Welp. As a blind person who has been looking forward to self-driving vehicles since the 2004 DARPA grand challenge, this is a not totally unexpected kick in the balls.

Most driving is fairly routine, but without strong AI I don't see how a self driving car can handle things like:

There is a car accident on the side of the road, and there is a human cop waving cars one at a time to drive past on the wrong side of the road.

I am part of a funeral procession, or a police escorted parade and I need to drive through red lights in order to stay in the procession. (That at least I know from the start to take manual control - except google wants a car with no wheel, and in a world with lots of such cars people won't know how to drive.)

Or, there is an ambulance behind me sirens wailing and in order to get out of the way I must drive illegally for a moment (through a red light, or on the wrong side of the road).

How about something as simple as driving the car up some ramps so I can service the car. How would you do that without both a steering wheel and experience in how to drive?

I don't see how a self driving car can handle any of those.

Self-driving cars should only need to handle 90% of driving tasks. For the rest, a qualified human can and should be required to take over.

That leaves humans without enough experience to actually know how to drive. It takes years of driving to get good at it.

If people drove only 1/10 as much as now no one would ever get good at it.

It's called "The Paradox of Automation" and it's a huge problem.

Basically the better the Automation the less often humans do anything, but it means that when humans do have to do something it's much much more critical than usual. Which is exactly the situations the humans have no experience with since they take control so rarely.

Basically some automation is good, but a lot of automation is actually detrimental to safety. For each industry you have to calculate that cutoff point and do no further automation past that point, until you are able to reach 100% perfection.

Maybe the cars should be required to pass a CA driver's test? If it's good enough to test a driver, why not a car?

This reminds me of the recent discussion about Baidu's expulsion from an image recognition competition. They submitted too frequently, causing their machine learning algorithm to solve specifically for the test, rather than for general purpose image recognition.

In your proposed scenario, knowing the CA driver's test constraints, the engineers could easily (and perhaps even inadvertently) be able to design a driver AI capable of passing the test with flying colors, yet fail horribly once faced with real-world driving.

That's not to say that's actually what would happen, only to say that the test would actually be rather useless to certify a self-driving car.

Another win for Test Driven Development!

A bad actor could in the future simply engineer the cars around the driver's test, and not the real world. In the future, it won't just be Google / Apple / Tesla making cars.

I'd hope that is the MINIMUM they would be required to pass.

, servant walking 20 paces in front carrying red flag.

Reply on news.ycombinator.com