That "Cruise" company worries me. It may give self-driving cars a bad reputation.
It seems to be lane-keeping plus radar cruise control, plus hype. All the major auto manufacturers have lane-keeping and cruise control working, and some are shipping it. It's good enough to work in most situations on freeways. Most. The auto manufacturers are reluctant to let customers use it as a full autopilot. The basic technology doesn't have enough situational awareness.
(Turn up the video resolution and read the displays. Messages include "lateral effort from planning too high", "Timeout from ... radar", "Planning timeout ... steer", "Curvature from planning too high".)
Cruise is hyping their system as a full driver-can-do-something-else autopilot. The major manufacturers, all of whom are painfully familiar with liability claims, aren't ready to go that far. Their systems have features which force the human driver to stay reasonably alert. Cruise is app and web people, not real-time and avionics people. This is worrisome.
Of course. I should clarify that I fully support self-driving cars. I've done research in the area, and I'm a big believer in their promise.
In fact, that's why I'm so aghast at these guys. They're (at best) delivering a 50% solution that will require human intervention. But their marketing strongly implies that you will be able to divert your attention from the road. A single bad self-driving implementation will poison the well for many years.
Lane following is a recipe for disaster.
There are many places in the Bay Area where lanes are loosely defined as "how many cars will fit". And then there are the variable lanes of GGB...
What happens when it reaches the end of the drivable highway? Will it wake me up and make me take the wheel? What if I don't? Does the car crash, emergency stop in the middle of the highway, or pull over gracefully?
There are many places you couldn't conceivably pull over.
You can always get around this kinds of problems by stating the functionality clearly (and limiting it's actions to a well defined set), but exactly this diminishes greatly the functionality of the service.
If you guarantee to people they can safely idle, they will idle and barely look at the road. If you don't, they will have to maintain attention at all times and all this will be is a "strain-reduction device" (i.e. not so useful). There's no good half-attention for humans.
It sucks when press (or even our own marketing) give well-meaning people like you the wrong impression. We need to fix this. The positive impact self-driving cars will have on society far outweighs any desire we have to be first to market or turn a quick profit. We haven't launched a product yet, and we won't until it's safe.
For example, we built a mashup of lane-keeping assist and adaptive cruise control within the first 3 months of starting the company. It wasn't safe enough, so we didn't launch it.
We're on roughly our 5th major iteration of the product. It uses maps, but it also works fine without them. It uses a GPS, but it works fine without it. It uses lane markers for localization, but it works fine without those too. But it's still not safe enough, so we haven't launched it yet.
Plenty of tough problems left to solve, but enough upside to justify our commitment.
The part that should make you feel better is where they knew it wasn't safe, so they didn't release it. Prototyping by mashing things together is how a significant amount of products get started. Including cars.
According to Wikipedia, a mashup is a term used to describe web development and combining music. It doesn't denote to me rigorous engineering development. And no, it doesn't inspire confidence in me that they didn't release a mashup that was unsafe.
I thought it was an odd adjective to use for a process where bugs equal death. He made a lot of money so that means a lowly developer can't question his choice of words ?
"This is the Cruise RP1, the first autopilot for your car. ... Frees you from the pain of commuting, stop and go traffic, and long trips. Just sit back and enjoy the ride. ... Reserve yours at getcruise.com and join the driverless revolution." - Cruise RP1 commercial: https://www.youtube.com/watch?v=uEQqQj_zyHI
"Your Personal Chauffeur Keep your car. Just let it drive you."
- getcruise.com
"Be the first to experience Cruise. We're currently accepting 50 pre-orders for a 2015 launch. Order now to join the driverless revolution."
- getcruise.com, a few months ago. (Now the site offers only a wait list.)
If there's a "wrong impression" of Cruise, it's because that was the impression they deliberately gave.
I have deep misgivings about almost-automatic driving. Full automatic driving seems to be within reach, and is not that far away. It's almost-automatic driving that's troubling.
Suddenly handing control back to the driver when the automatic system can't cope is not going to work well. Commercial pilots have problems with that. Watch "Children of the Magenta" (https://www.youtube.com/watch?v=pN41LvuSz10), where a chief pilot of American Airlines talks to other experienced pilots about how to use cockpit automation. Those are very well trained pilots. Yet this talk is about accidents which usually involved the end of a period where automated systems were flying the airplane, and then, for some reason, from conflicting traffic to weather to a runway change, the pilot had to take over. At that moment, sometimes the pilot didn't have full situational awareness, because they weren't flying the aircraft. That's when accidents happened.
Drivers are not selected and trained like airline pilots. An almost-autopilot puts them in a situation where they may have to take over suddenly. One cannot expect drivers to do that with a high degree of success.
The minimum safety requirement for automatic driving is probably to be able to pull over and come to a complete stop after a failure without user intervention. Or to drive slowly enough that slamming on the brakes is an adequate solution. That's probably why Google wants to initially deploy automatic driving on vehicles with a 25MPH top speed. At that speed, an emergency stop is a sufficient response to almost all problems. There's no need to be able to drive out of trouble at 25MPH.
A high level of driver ineptitude has to be well tolerated. People are going to use automatic driving when tired, drunk, stoned, or looking at a phone. That has to work, and work well. If not, people will die. This is a very tough human factors problem.
It would be terrible if a driver erroneously thought that the Cruise system was engaged when it wasn't, and crashed because of that. How do you try to minimize that risk?
I'm all for self-driving cars, but what I want to hear about is how on earth they'll overcome things like inclement weather, pedestrians, poorly maintained roads, and combinations thereof... but maybe that's just because I'm currently in New England in the winter.
Bad weather and unexpected obstacles like pedestrians are actually trivial to handle. Remember that an autopilot's "visual" perception isn't limited to the visible-light range like ours is.
Poorly-maintained roads are the problem that would give me nightmares if I worked on these. Lane markers that might have been perfectly usable when the database was compiled may be too worn out or obscured by snow, leaves, or who-knows-what else to be usable later. How can you design a system to fail safely when every input source is independently fallible? And how can we redesign the auto insurance industry to make it possible to try?
Detection of abnormal conditions might be tractable (though I wouldn't call online vision or recognition trivial problems). But taking action is harder. A piece of debris might manifest in a way that causes a self-driving car to skid to a halt or swerve - dealing with abnormal conditions in real time on crowded roads is certainly not a trivial problem, especially given the stakes.
When can we have an intelligent discussion about what licenses for self driving cars would look like?
We license drivers now. It seems clear that software should have to pass at least the same bar.
Unfortunately everytime it is raised it ends up being a discussion on the effectiveness (or lack thereof) of regulation.
My view is that Google's version of a self driving car may be perfect, but I'm unconvinced that "vendor modifications" (to steal a term from the Android world) will always be as well-coded.
What would an appropriate regulation/certification/licensing regime look like? How are updates handled? Are users allowed to modify the software themselves (with or without re-certifying?)
Even Google's version will have risks and I expect will make mistakes in early iterations. Having a max speed of 25 MPH and limiting deployment will reduce some of this risk, but AI is hard and the environment is very open and will constantly include novel states the agent has to make sense of. Personally I'm concerned the initial Google design doesn't include a steering wheel for manual override - seems overconfident to me even for normal scenarios let alone degenerate ones (weather, internet connectivity, GPS malfunction, construction, bad maps, road damage, etc.).
I would vote, that the self-driving car revolution started more than 20 years back, when Mercedes, Bosch and others already worked on the self-driving cars. The times when there where no such thing as GPS available to the general public. All the assistants we have these days in cars like, auto distance control, drive lane assistant, obstacle detection, stop-and-go assistant and so on evolved from those prototypes.
At that time a made an internship at the military arm and a fellow had an internship with the self-driving car.
BTW, the most current Mercedes S-Class can do self-driving for about 80% of the time. They intentionally disabled it and even put in logic to detect if the driver is holding the steering wheel or not. If the driver is not holding it for some time, the assistants are disabled by intention.
Also what I do not understand, why Cruise is attaching everything mechanically? Any descend car these days with lane keeping assistant/self-parking assistant has all the parts available. You just have to connect to CAN and control the whole car. Of course, you have to talk with the OEMs. But that is how the companies are work, which convert cars for use for people with disabilities.
32 comments
[ 4.8 ms ] story [ 72.2 ms ] threadIt seems to be lane-keeping plus radar cruise control, plus hype. All the major auto manufacturers have lane-keeping and cruise control working, and some are shipping it. It's good enough to work in most situations on freeways. Most. The auto manufacturers are reluctant to let customers use it as a full autopilot. The basic technology doesn't have enough situational awareness.
Here's what Mercedes has right now:
https://www.youtube.com/watch?v=r_RFzC_G5BA
(Turn up the video resolution and read the displays. Messages include "lateral effort from planning too high", "Timeout from ... radar", "Planning timeout ... steer", "Curvature from planning too high".)
Cruise is hyping their system as a full driver-can-do-something-else autopilot. The major manufacturers, all of whom are painfully familiar with liability claims, aren't ready to go that far. Their systems have features which force the human driver to stay reasonably alert. Cruise is app and web people, not real-time and avionics people. This is worrisome.
In autonomous driving marketing, Volvo is the current leader: https://www.youtube.com/watch?v=42E-tF-6PWU
The agile startup mentality is great until you try to apply it to domains that are literally life-and-death.
http://wfdd.org/post/why-we-judge-algorithmic-mistakes-more-...
In fact, that's why I'm so aghast at these guys. They're (at best) delivering a 50% solution that will require human intervention. But their marketing strongly implies that you will be able to divert your attention from the road. A single bad self-driving implementation will poison the well for many years.
You can always get around this kinds of problems by stating the functionality clearly (and limiting it's actions to a well defined set), but exactly this diminishes greatly the functionality of the service.
If you guarantee to people they can safely idle, they will idle and barely look at the road. If you don't, they will have to maintain attention at all times and all this will be is a "strain-reduction device" (i.e. not so useful). There's no good half-attention for humans.
It sucks when press (or even our own marketing) give well-meaning people like you the wrong impression. We need to fix this. The positive impact self-driving cars will have on society far outweighs any desire we have to be first to market or turn a quick profit. We haven't launched a product yet, and we won't until it's safe.
For example, we built a mashup of lane-keeping assist and adaptive cruise control within the first 3 months of starting the company. It wasn't safe enough, so we didn't launch it.
We're on roughly our 5th major iteration of the product. It uses maps, but it also works fine without them. It uses a GPS, but it works fine without it. It uses lane markers for localization, but it works fine without those too. But it's still not safe enough, so we haven't launched it yet.
Plenty of tough problems left to solve, but enough upside to justify our commitment.
"Your Personal Chauffeur Keep your car. Just let it drive you." - getcruise.com
"Be the first to experience Cruise. We're currently accepting 50 pre-orders for a 2015 launch. Order now to join the driverless revolution." - getcruise.com, a few months ago. (Now the site offers only a wait list.)
If there's a "wrong impression" of Cruise, it's because that was the impression they deliberately gave.
I have deep misgivings about almost-automatic driving. Full automatic driving seems to be within reach, and is not that far away. It's almost-automatic driving that's troubling.
Suddenly handing control back to the driver when the automatic system can't cope is not going to work well. Commercial pilots have problems with that. Watch "Children of the Magenta" (https://www.youtube.com/watch?v=pN41LvuSz10), where a chief pilot of American Airlines talks to other experienced pilots about how to use cockpit automation. Those are very well trained pilots. Yet this talk is about accidents which usually involved the end of a period where automated systems were flying the airplane, and then, for some reason, from conflicting traffic to weather to a runway change, the pilot had to take over. At that moment, sometimes the pilot didn't have full situational awareness, because they weren't flying the aircraft. That's when accidents happened.
Drivers are not selected and trained like airline pilots. An almost-autopilot puts them in a situation where they may have to take over suddenly. One cannot expect drivers to do that with a high degree of success.
The minimum safety requirement for automatic driving is probably to be able to pull over and come to a complete stop after a failure without user intervention. Or to drive slowly enough that slamming on the brakes is an adequate solution. That's probably why Google wants to initially deploy automatic driving on vehicles with a 25MPH top speed. At that speed, an emergency stop is a sufficient response to almost all problems. There's no need to be able to drive out of trouble at 25MPH.
A high level of driver ineptitude has to be well tolerated. People are going to use automatic driving when tired, drunk, stoned, or looking at a phone. That has to work, and work well. If not, people will die. This is a very tough human factors problem.
Poorly-maintained roads are the problem that would give me nightmares if I worked on these. Lane markers that might have been perfectly usable when the database was compiled may be too worn out or obscured by snow, leaves, or who-knows-what else to be usable later. How can you design a system to fail safely when every input source is independently fallible? And how can we redesign the auto insurance industry to make it possible to try?
Google's self-driving cars still have trouble with heavy rain and snow (or at least they did as of last year). http://www.technologyreview.com/news/530276/hidden-obstacles...
I'd be curious to hear what the insurance industry makes of all of this too, they must be preparing for it in some way or another.
We license drivers now. It seems clear that software should have to pass at least the same bar.
Unfortunately everytime it is raised it ends up being a discussion on the effectiveness (or lack thereof) of regulation.
My view is that Google's version of a self driving car may be perfect, but I'm unconvinced that "vendor modifications" (to steal a term from the Android world) will always be as well-coded.
What would an appropriate regulation/certification/licensing regime look like? How are updates handled? Are users allowed to modify the software themselves (with or without re-certifying?)
At that time a made an internship at the military arm and a fellow had an internship with the self-driving car.
BTW, the most current Mercedes S-Class can do self-driving for about 80% of the time. They intentionally disabled it and even put in logic to detect if the driver is holding the steering wheel or not. If the driver is not holding it for some time, the assistants are disabled by intention.
Also what I do not understand, why Cruise is attaching everything mechanically? Any descend car these days with lane keeping assistant/self-parking assistant has all the parts available. You just have to connect to CAN and control the whole car. Of course, you have to talk with the OEMs. But that is how the companies are work, which convert cars for use for people with disabilities.