Google's head of self-driving cars' case against driver assistance (AutoPilot)

[Vger]

I found this TED talk quite interesting, putting what Tesla is developing in context. While I do not agree with the full argument (that driver assistance cannot asymptotically become self driving), seeing something of the innards of how much they have developed over the past three years gives a whole new sense of just how difficult the total problem is:

Chris Urmson: How a driverless car sees the road | TED Talk | TED.com

 

[MsElectric]

I found this TED talk quite interesting, putting what Tesla is developing in context. While I do not agree with the full argument (that driver assistance cannot asymptotically become self driving), seeing something of the innards of how much they have developed over the past three years gives a whole new sense of just how difficult the total problem is:

Chris Urmson: How a driverless car sees the road | TED Talk | TED.com

Thanks so much for sharing. It is great to see how sophisticated and thorough the Google self driving algorithms are becoming. I too don't agree with everything he said but when watching the video the biggest thing I realized is the importance of a comprehensive sensor suite.

I really hope the Autopilot Ver. 2.0 sensor suite includes pretty much 360 radar and more than that they have LIDAR sensors around the car for overlapping and redundant coverage. What makes the Google self driving technology possible is having all that great LIDAR data around the car. With that it is then just a matter of software.

 

[ww73]

Here's an interesting podcast that explains the reasoning for Google's "UI". I'm not saying I agree, but it's an extra data point.

Episode 642: The Big Red Button : Planet Money : NPR

 

[brkaus]

Thanks so much for sharing. It is great to see how sophisticated and thorough the Google self driving algorithms are becoming. I too don't agree with everything he said but when watching the video the biggest thing I realized is the importance of a comprehensive sensor suite.

I really hope the Autopilot Ver. 2.0 sensor suite includes pretty much 360 radar and more than that they have LIDAR sensors around the car for overlapping and redundant coverage. What makes the Google self driving technology possible is having all that great LIDAR data around the car. With that it is then just a matter of software.

With redundant radar and LIDAR it isn't a matter of software, it is a matter of cost. LIDAR is quite expensive from my understanding.

 

[ItsNotAboutTheMoney]

I agree that the current approach has limitations, but I disagree that it doesn't help get to autonomy. Better recognition leads to better assistance, and therefore I'd expect a gradual increase in the capability of assistance systems.

 

[cryptyk]

I'm really excited about auto-pilot and driverless cars. I'm a software engineer, and a technologist. I'm an early adopter, and I'm really optimistic about technology. All of that said, I don't think we'll see self driving cars anytime soon. It has nothing to do with the technology; it's just society and policy.

I don't think we'll ever solve the Trolley Problem. Imagine a scenario where your self driving car detects a 5 year old girl running across the street. Your car has enough data and time to make a choice: run over the girl, killing her, or smash into a barricade, killing the occupants of the car. How should we program the car in this case, and in all of the hundreds of cases like it? How do we determine the relative value of people's lives?
What if there are four people in the car? Does that change the answer?
What if the person crossing the street is a 94 year-old homeless man?

When we make snap decisions as humans, the outcome can be called tragic, unfortunate, and "an accident".
When we program our cars to decide in these situations, it ceases to be "accidental". Now, it's calculated, reviewable, and intentional.

I wouldn't want to be the guy to explain to a mother that her child was killed because the car calculated that her life was less valuable than the people in the car.

 

[Johan]

I'm really excited about auto-pilot and driverless cars. I'm a software engineer, and a technologist. I'm an early adopter, and I'm really optimistic about technology. All of that said, I don't think we'll see self driving cars anytime soon. It has nothing to do with the technology; it's just society and policy.

I don't think we'll ever solve the Trolley Problem. Imagine a scenario where your self driving car detects a 5 year old girl running across the street. Your car has enough data and time to make a choice: run over the girl, killing her, or smash into a barricade, killing the occupants of the car. How should we program the car in this case, and in all of the hundreds of cases like it? How do we determine the relative value of people's lives?
What if there are four people in the car? Does that change the answer?
What if the person crossing the street is a 94 year-old homeless man?

When we make snap decisions as humans, the outcome can be called tragic, unfortunate, and "an accident".
When we program our cars to decide in these situations, it ceases to be "accidental". Now, it's calculated, reviewable, and intentional.

I wouldn't want to be the guy to explain to a mother that her child was killed because the car calculated that her life was less valuable than the people in the car.

This is overthinking it. Humans cause collisions and autonomous vehicles will cause collisions. It won't be as controversial as some seem to think. The car should follow the traffic rules as best it can, and will likely drive in to the girl given your example, whilst braking. The car does its best to slow down but it's an impossible situation. The girls parents shouldn't have let her run in to the road.

It's uncontroversial really.

When humans collide it's no less or more intentional than if an autonomous car does.

 

[cryptyk]

This is overthinking it. Humans cause collisions and autonomous vehicles will cause collisions. It won't be as controversial as some seem to think. The car should follow the traffic rules as best it can, and will likely drive in to the girl given your example, whilst braking. The car does its best to slow down but it's an impossible situation. The girls parents shouldn't have let her run in to the road.

It's uncontroversial really.

When humans collide it's no less or more intentional than if an autonomous car does.

I think the difference is that we can decide ahead of time once we have sophisticated autonomous cars. You're choosing to run the girl over. Others might decide it's better to hit the barricade. In any case, we get to decide how the car will act while we sit in our living rooms, instead of the heat of the moment.

 

[Cosmacelf]

Yes, but at the end of the day, the car must be programmed to save the occupants. Not to mention that the software doesn't actually know with 100% certainty it is a little girl. It could be a doll someone threw out there, or a ballon in a human shape.

The trolley problem won't the reason we don't have autonomous vehicles...

 

[Johan]

I think the difference is that we can decide ahead of time once we have sophisticated autonomous cars. You're choosing to run the girl over. Others might decide it's better to hit the barricade. In any case, we get to decide how the car will act while we sit in our living rooms, instead of the heat of the moment.

That's not how advanced semi-intelligent software works. You don't pre-program the response to every single possible situation that can arise. The car makes the decision, not a human programmer beforehand, hence "autonomous".

 

[Cosmacelf]

Ok, so I just listened to the Google TED talk. Thanks for posting it!

How can I say this with tact? Guys, this was a presentation done by an expert who deeply understands the problem at hand. He lives and breathes this stuff every single day, and has millions of miles of data to backup his assertions. And people here blithely say he's wrong? "Oh well, I just don't believe this expert, I'm going with my gut and say that driver assistance can get us to driverless cars". Based on what? Wishful thinking?

I totally believe this guy. Actually I was pretty impressed that they are completely recording (with all telemetry channels) every single inch of driverless car mileage which allows them to make the system very smart indeed. I wish Tesla were half as smart - Tesla could fix their frequently brain dead navigation system if they learned the actual routes people took to get to destinations.

On this particular issue (autonomous driving) Tesla has a half baked solution that will not live up to Elon's hype. Google and Google alone appears to have driverless technology and if they license it to someone other than Tesla, you can kiss Tesla the company goodbye. Electric cars are cool, but driverless cars are cooler.

 

[MarkS22]

I think the difference is that we can decide ahead of time once we have sophisticated autonomous cars. You're choosing to run the girl over. Others might decide it's better to hit the barricade. In any case, we get to decide how the car will act while we sit in our living rooms, instead of the heat of the moment.

An autonomous car will do everything to avoid an accident. In an extreme fringe case, like a child running into the street, you're talking about: (1) The child was completely hidden behind something (and hidden to considerably more sensors than human eyes, including multiple cameras, radar, LIDAR, etc) and (2) the child darted into a roadway where the car was going the speed limit, but without time to stop with a near-perfect reaction time and expert defensive driving techniques.

Remember, when going 30mph, alert humans waste roughly 50ft for reaction time and 30ft for moving their foot from accelerator to brake after deciding to take action. Once the brakes are engaged, the car will stop in, say, another 40ft. An autonomous car will react and brake in 40ft. So, they're stopping 80ft sooner at a relatively slow 30mph. (The difference is more dramatic as speeds increase.) I think people underestimate the huge reaction time (mental and physical) that humans add to actually avoiding an accident. To me, this nullifies the argument of calculation on "value of life" because--given identical scenarios--the autonomous car has stopped before a human could even make a decision and engage the brakes or swerve.

 

[Cosmacelf]

Actually I just realized what Apple is doing with their rumored electric car project. They will likely license googles tech and have a purpose built electric driverless car. This would appeal to Apple since it is a green field design problem - what does a car look like that has no steering wheel and no driver? And the rumored 2020 date would allow Google to perfect their technology before Apple used it.

Meanwhile Tesla has no real program in place for a truly autonomous vehicle.

 

[MarkS22]

Meanwhile Tesla has no real program in place for a truly autonomous vehicle.

MobileEye is the providing at least some of the Autopilot hardware for Tesla. They had their own presentation earlier this year demonstrating their technology and road-map to a fully autonomous car, specifically addressing how they differ from Google: The Future of Computer Vision and Automated Driving by Prof. Amnon Shashua - YouTube

It's particularly interesting to compare what their systems "see" in comparison to Google. To me, it seems as though Google is downplaying the ability for driver assist to evolve because they're attacking the problem from a different direction. Maybe they're right, but MobileEye isn't a tiny company... they're valued at 12 billion and dedicated strictly to optical image systems for the auto industry. Sure, Google is 37 times larger, but how much of that is dedicated to autonomous driving?

 

[Matteo]

I have to say I agree with the Google guy. Tesla's radar can see things in front of the car, however it is not a smart system that recognises a cyclist or pedestrian and then anticipates future movements.

However, Google's system does. If you watched the video, their system recognizes a pedestrian, a cyclist, traffic cones etc and then reasons based on that information to think what might happen next. For example it keeps watching the cyclist, anticipating that it might ignore the traffic lights. Another amazing example was when it saw construction work ahead and anticipated that the car in front would change lanes. That's very smart.

Tesla's system is not smart. It can't differentiate a wall from a parked car. You can see how it fails in this video. Because the blue car moves, it recognizes this as a vehicle. But because the white car doesn't move, it sees that as any obstacle like a wall, parked car, roadside barrier etc. It assumes traffic flows with the blue car and you will just follow that.

There is also this other issue about asking the driver to pay attention at all times. That wont work. Again I agree with Google's approach here since I read the following quote last year:

Source: MAY 27, 2014 NYTimes article

 

[S'toon]

http://xkcd.com/1559/

 

[Electroman]

On the situation of a car running over a darting kid - give that the autonomous car can react much quicker than human, an accident by autonomous car is much more easily defensible compared to a human driver in those same conditions.

Dontcha think?

 

[Johan]

On the situation of a car running over a darting kid - give that the autonomous car can react much quicker than human, an accident by autonomous car is much more easily defensible compared to a human driver in those same conditions.

Dontcha think?

Exactly. You can even prove in retrospect that within the laws if physics the accident was unavoidable. So there should be less hard feelings and blame to be placed. Sort of like how if an earthquake or tsunami takes a life it's very difficult to get mad at the tsunami. These things sometimes happen. We can design the cars, roads, surroundings to our best abilities. And we can watch our children. But there will likely be accidents in the future, however much fewer (just like has been the case in the last decades: more traffic, fewer accidents).

 

[Electroman]

A rough analogy are the accidents that happen on rail crossings. The train is never at fault.

 

[WarpedOne]

I have to say I agree with the Google guy. Tesla's radar can see things in front of the car, however it is not a smart system that recognises a cyclist or pedestrian and then anticipates future movements.
However, Google's system does.

Google guys premise was that pimping up driver assistance systems will lead to problems and cannot lead to self driving car.
Which is pure marketing BS.

What google's system does now and what tesla's system does now cannot be compared without equalizing the price.
The whole point of throwing away lidar is its cost. Not its problems or incapablity, but its cost.

Tesla's current suit of sensors costs $2500. They will probably have to add some more and increase price say to $5000.
What is the added cost of google's system? 50.000? 150.000?

And at the end of the day tesla still drives the speed limit. Google self driving drives at 25?