So the 3d radar used by Google costs $75,000, but it seems unlikely they'd try to sell the tech to the mass market at that price. How likely is it that this is just the price of their R&D units, and that they'll have that reduced significantly for production models? What do you HN-ers think is a reasonable price for this kind of technology?
In any case, congrats to the student! His work sounds impressive.
I think you are right about it being R&D cost. I'm sure its more than a webcam even at scale but I think that both probably have their pros and cons. How do the cams work the morning after a snowstorm when everything is white vs the radar system? I could see both being in-use with one being used as warm backup based on road conditions. Never the less still sounds like the guy knows his stuff.
I think that the first Google car will be very expensive anyhow. One of the reasons is, that you can fit a $40k sensor package into a $400k car. And going with a similar argument, self driving is probably a feature that warrants a 10% - 30% markup, so for a $40k car somewhere between $5k and $15k.
That might very well be the average markup the general public is willing to pay, but don't discount the niches that will tolerate a much higher premium - these will provide the initial traction.
Someone that is legally blind might gladly pay an additional $40k on top of a $30k car. So might someone with a two hour commute, or that likes to get drunk a lot.
How much is it worth to have police cars that can drive themselves back when an officer is injured? Maybe the asking price, maybe not. But you can be sure it will be considered.
Segmenting the market based on need, versus average car spend, seems to me a much better way to look at it.
What about LIDARs that aren't mounted on cars themselves but on streets which all cars can connect to. This will limit the useful range of such cars, but imagine manhattan decked out in these things and taxis or busses that only work in Manhattan. Could still have huge savings, especially if combined with computer vision as a fallback.
Edit: To be honestly my idea's probably not worth pursuing. I fully expect the cost of LIDARs to come down drastically in the next five years.
That would probably save money, but it'd be one of those large cash amount up front initiatives that tax payers wouldn't agree to until every other car already self drives.
IMO, there is some market for private self driving cars at almost any price, if it's usable and legal. The minimum volume at which they can be produced, legalized may be too high though for any kind of niche luxury market though.
The path to substantially replacing human driven cars though is (I think) some sort of driverless taxis. The price where driverless taxis take over is probably somewhere around where the cost per km of driverless is lower. That could be pretty high. Assuming other operational costs are similar to regular taxis $150k-$200k for a driverless vehicle that lasts 3-5 years sounds viable.
Driverless cars have been sitting around the edge of our consciousness as "experimental but interesting" for a long time. But, the economic changes they introduce are potentially enormous, wide. I think there's a good chance that privately owned cars become rare, for example. General purpose cars may go away. A golf cart is just fine for going 5-10 miles in urban traffic.
This space is big enough for Google/Ford/IBM sized giants to emerge in. Makers of the cars, makers of the driver, makers of the UI, operators of the taxi companies, who knows what else. Maybe makers of digital roadsigns. Maybe selecting destinations from a screen creates the kind of effect Adwords has had.
The thing that makes me think of taxis is the the driver cost/total cost ratio is high, so removing the driver is a big deal.
Buses and long haul trucks are already pretty highly optimized in that the driver's time is a relatively smaller component of the total cost. Also, the total cost is much lower.
> A golf cart is just fine for going 5-10 miles in urban traffic.
I hadn't thought of this. Imagine Google buying garages all over NYC and filling them with driverless Smart cars (or similar). A button is added to your phone's Google Maps app, so that in addition to "Get Directions" and "Navigate", you also have the option of "Pick Me Up".
So, you search for where you want to go on Google Maps and tap "Pick Me Up", with the option of telling it how many people are in your party. Within 5 minutes enough Smart cars for your party stop within 20 feet of you, even if you've already started walking, thanks to your phone's GPS. You get in the car, which has no steering wheel, and touch "Go" on the in-dash touch-screen to start the car driving to your destination.
Once again the title of a post doesn't remotely represent its contents. In this case a student's solution costing $71,000 less passed 47 of 50 self imposed trials.
Google on the other hand have driven half a million miles with their solution.
Object recognition from video is probably the future, it's how we work after all. The title though? Quite misleading.
3/50 is an unacceptably high failure rate for automobiles... "in-factory" failure rates are measured by PPM (parts per million) and in-field failures I imagine must be in the order of "trillionths" in order to maintain reliability across millions of cars driving tens of thousands of miles per year each.
edit: I don't mean to dismiss the student's accomplishments, but wanted to bring some perspective
What he did was utilize an additional channel of input, and combine the two channels. Adding this kind of additional input system might actually enhance safety, as a failed radar component would still leave the vehicle with partial sensory capacity. (Also: limited capabilities of the current model could easily be effectively reduced via an enforced speed limit or similar. Failures were at 20-30m, which is a serious distance at average driving speeds in urban environments.)
It's not because they are measured by PPM that they are not high :P I have seen cases where a certain process deviates and gives more than 10 000 PPM of "out of range" quality. What's important is to understand the impact of the deviation and if this has a negative impact on the finished product quality. In short the answer is "not always".
To mitigate the issues in complex systems like planes and so on, the systems are built with multiple redundancies in place. That way even if something fails, the system remains stable and operative. So you don't need to have a "per trillionth" (and you are using the wrong word) quality system in order to deliver a very robust system.
I think the student is perfectly aware that his product must be improved a lot before hitting the road. But it's a starting point. Right now it is 94% accurate in simulations, the next version might be 99%. I'm sure in time if he can get a team together to work with him and continues to innovate, he will make it into a product that meets the industry standards.
3/50 is I think a meaningless failure rate. If the system overlooks 3 out of every 50 pedestrians standing on the road by themselves and happily runs them over, then it is certainly unacceptable. If there is a test, in which the system is shown groups of pedestrians, in total 50 and it mistakes a group of 3 for a large immobile object, then it is probably acceptable, as long as it always labels pedestrians as 'do not run over.'
And for the entire fleet of autonomous cars, the system does not need to be safe. It just needs to be better than humans driving. Realistically to gain acceptance they would need to be an order of magnitude better, which translates to roughly 1 fatal accident per year per 10^5 self driving cars. The current numbers for the US are 36000 fatalities and 250M cars.
> And for the entire fleet of autonomous cars, the system does not need to be safe. It just needs to be better than humans driving.
I really hope this is true. I really want self-driving cars everywhere, now. Road Traffic accidents kill very many people and this is an example of Google using money and smarts to do good.
But looking at the way people deal with risk makes me wary.
People drive cars all the time even though driving is a bit risky. People don't really understand how good or bad their own driving is. There's a bunch of cognitive biases and rationalisations.
I hope people who are experts in communications are ready to dispel the FUD backlash against self driving cars.
> But looking at the way people deal with risk makes me wary.
One of the problems is insurance. Right now, everyone needs to have insurance to drive. If I accidentally hit a pedestrian or rear-end another driver they can claim on my insurance.
Let's say that a software bug results in someone getting hit and injured by a self-driving car. Who's liable? I am not sure that this is FUD, so much as a genuine question. It is because driving is risky that drivers must have insurance. Presumably most people would want self-driving cars to be insured if there was a chance of failure (even planes, one of the safest forms of transport, get insured against crashing).
We could very well have a situation where cars self drive, but drivers are still expected to pay attention to the road and 'drive' to avoid liability.
I suspect there's going to be a lot of lobbying on both sides before mass deployment of self driving cars.
Insurance is an industry very much in need of disruption. If Google offered a monthly software license for a self driving car that also covered the insurance? That would be awesome, and many people would probably take them up on it.
I think we have quite some experience with insuring unsupervised systems. After all, I am not supposed to supervise my coffee machine, but if it causes a fire then my insurance has to pay.
On the other hand, I think that the politics of self driving cars could get rather interesting, especially if widespread adoption is fast, say 5 years until a self driving car is no longer extraordinary and 10 years until one just assumes that a new car is self driving. Then I think we get a initial phase, were they are seen as modern. After this, when every single accident gets widely reported, I think it is possible that the public completely splits, on one hand the traditionalists, who maintain that only a human should drive. And on the other hand the people who say that self driving cars are safer than humans. And the laws will shift according to the relative strength of the two groups. ( With probably some rather ridiculous political compromises in between, perhaps you are allowed to read a newspaper, but not to drink coffee. )
This is, at best, a first-pass alpha, but a highly promising one. Further, the student apparently understood the reason for the errors and offered an inexpensive way to fix them. What would you have wanted int the title?
"Student takes Google's self-driving car and offers the potential to make it $71,000 cheaper contingent upon correcting a misidentification issue that causes a possible failure rate of 6%"
Yeah, that would be so much better ;)
It's OK to take a little license with a title and try to summarize the main points. Maybe just adding the word "potential" in there would be enough, but some user named lololinternet would come along and claim how misleading the title is because ...
He didn't "take" their car and do anything with it. He made a lower cost version of one of the systems, but not a fully functioning car. This is hugely impressive, but that doesn't mean the title isn't misleading.
Has anyone located a more detailed project description or paper? Using camera imagery is definitely nothing new, and in many autonomous systems with radar/lidar (including the Google car), cameras still supplement the lidar imagery. I'm wondering how he improved on the current research.
Everytime I see a news report from a field that I'm familiar with, I get disappointed...
Definitely, congrats to the student! He really did astonishing work. But, nothing that could be useful in production. As I'm familiar with use of computer vision in traffic from academic and industrial point of view, I know situations this system has to deal with. And, I know what are the state-of-art results in that area.
Computer vision is heavily used in traffic, but, self driving car is still out of its reach.
There's also a lot of work going on with driverless technology.
Disclaimer: I work for a Conti subsidiary that develops camera-based surround view and object detection systems but not directly on the products (IT Support)
I don't have time for a complete post (and I don't have permission to publish examples from the datasets), but here is some "quick" reply.
Once you see examples from the datasets, lots of problems come to mind (and to todo list a bit later). This is a quite expensive problem to tackle with.
Some requirements:
(1) you need datasets from various places, various weather conditions and various situations,
(2) everything must work in realtime,
(3) equipment is expensive (cameras, cars, gas, ...),
(4) error has to be minimal (we are talking about human lifes).
(and this is just part of requirements)
Student labs fail at money part, companies fail at lack of time (again, money; you need lots of time to deal with extreme number of situations and produce error prune product -- and nobody guarantees that you'll manage to do that).
And now, some problems:
- Everything has to work in realtime which means more than 30 FPS in average (you need to aim for higher average speed so you don't get lags in complex scenes). Computer vision algorithms aren't usually realtime, for example, for simple task as object detection one of the best realtime algorithms is Viola-Jones which is more than 10 years old and patented. In newer days there are some breakthroughs in this area, but quite small if you take ten-year gap (and we are talking about quite simple problem -- object detection).
- Then, the datasets. You need a lot of them. Different places, different times. All you can do is to beg someone for it, pay a lot for it or make an contract with traffic companies (for the product that you don't know will it work well; and good luck if you represent a student lab).
- Now, the data. Take a ride during the different weather conditions and times of day or year. You'll be ok, but from CV point of view, you'll meet hundreds different problems. Disorted view during the rain, big balls of light during the night, different types of cars (cars with trailers, bikes at the back, motorbikes, ...), damaged road, damaged traffic control, traffic accidents, ... you get the point.
- Then, think of number of things that you need to take care of -- traffic signs (very hard problem! specialy if you want to ride on local roads where some of signs are only partially visible), traffic fixes, etc.
Some of solutions are taking only some of the problems and tacking with them (more in way of alerting the driver). I'm not familiar with different sensors that deal with some of these problems, but maybe some of them aren't so expensive (today you can buy a car which can park itself (or that was just R&D showcase)).
Anyway, this is extremely interesting problem and it deserves us to fight with it. Unfortunately, from business side it looks like a big gamble.
It's a shame that the world of IT has got us to the point of assuming system crashes are normal for computer systems. We can build highly reliable platforms - they do cost more, but reliability is an understood challenge.
Does anyone know how he addressed the issue of bright lights hitting the camera sensor? or too low of light levels for he camera sensor to pick up quickly? I imagine that was the reason a $70,000 dollar 3d radar was even used in the first place.
I wonder how cheap every other car manufacturer that has been working on self-driving technology can make it. I also wonder if tech geeks even realize they have been, especially when Google's "self-driving cars" is cited as a reason for Google's continued growth.
The HN title is really bad and just plain misleading. It needs to be changed to the original news article title. The student didn't take a Google prototype and modify it. He built his own car and software. I'm pretty sure none of it was from Google.
From my vantage point the utility of self driving cars is in highway driving. I don't really care for auto city driving.
In that context perhaps the problem is somewhat simpler if one assumes the infrastructure will be modified to help these cars. A lot could be done if each car communicated with roadside beacons and other locating systems. Some will voice concerns about privacy. I couldn't care less. Your movements are already far more traceable than they were fifty or a hundred years ago.
If I could manually get on the Golden State; engage auto pilot and be alerted a couple of miles before my exit in San Jose I'd be thrilled. I wouldn't even care if the thing pinged me every 15 minutes to see if I am awake.
This doesn't help anyone's paranoia coming from robomartin! (Kidding aside, I think your profile cv got a little mangled. Try double line-breaks and log out for a sec so you can see it.)
Personally I would much rather have better above-ground transportation in the city than on the highway, but at that point I should be wishing for self-driving buses.
If we relinquish control of our automobiles to a central system, even busy intersections would not require vehicles to slow down or speed up more than ~5 mph to all safely pass through. Just one example of how auto-autos could achieve an unprecedented balance of safety and efficiency.
My question is, how far beneath the current human-driver accident rate will the AI-driver accident rate need to be before you accept that it is a better driver than you?
I think the language used to push these needs to be centered around protecting good drivers from all the other bad drivers. "You aren't the problem, they are the problem and these cars are simply able to better anticipate the mistakes others will make."
So you pair up this student with the high schooler who created the supercapacitor, and they'll disrupt a Google-Tesla alliance years before it's even conceived.
The article, like most articles uses an attention grabbing headline from the "$71,000 cheaper" point. The point of that is to give your brain a low resolution picture of the story. Enough to interest you and get you to read. Regardless of how you feel about that, if it worked (you are reading the article) it's best to forget about the conclusions/objections you started developing from the headline after you read the article. They're just a distraction.
Obviously, this isn't a drop in replacement for Google's system. Google probably haven't even been optimized their system for price anyway.
The interesting points here IMO is that (a)Google system cost ~$75k (b) a student was able to cheaply play in this space and get somewhere and (c) he did it by replacing the expensive 3d radar with a cheap 3d radar + webcam.
Insurance companies will charge a good bit less for insuring self driving cars, since they're already statistically safer than most human drivers. I imagine that that will help out adoption.
I am serious, they've already gone over a quarter million miles without any accidents that weren't someone else's fault. I'm not sure if that's better than average yet, but it seems like it's close if it's not already better.
60 comments
[ 2.9 ms ] story [ 116 ms ] threadIn any case, congrats to the student! His work sounds impressive.
Someone that is legally blind might gladly pay an additional $40k on top of a $30k car. So might someone with a two hour commute, or that likes to get drunk a lot.
How much is it worth to have police cars that can drive themselves back when an officer is injured? Maybe the asking price, maybe not. But you can be sure it will be considered.
Segmenting the market based on need, versus average car spend, seems to me a much better way to look at it.
Edit: To be honestly my idea's probably not worth pursuing. I fully expect the cost of LIDARs to come down drastically in the next five years.
The path to substantially replacing human driven cars though is (I think) some sort of driverless taxis. The price where driverless taxis take over is probably somewhere around where the cost per km of driverless is lower. That could be pretty high. Assuming other operational costs are similar to regular taxis $150k-$200k for a driverless vehicle that lasts 3-5 years sounds viable.
Driverless cars have been sitting around the edge of our consciousness as "experimental but interesting" for a long time. But, the economic changes they introduce are potentially enormous, wide. I think there's a good chance that privately owned cars become rare, for example. General purpose cars may go away. A golf cart is just fine for going 5-10 miles in urban traffic.
This space is big enough for Google/Ford/IBM sized giants to emerge in. Makers of the cars, makers of the driver, makers of the UI, operators of the taxi companies, who knows what else. Maybe makers of digital roadsigns. Maybe selecting destinations from a screen creates the kind of effect Adwords has had.
Stuff with fixed and repeating routes like buses (or airport shuttle buses) are also interesting imo
It would also be cool if down the road (pun not intended) a care could detect if the driver is drunk and switch to selfdrive mode.
Buses and long haul trucks are already pretty highly optimized in that the driver's time is a relatively smaller component of the total cost. Also, the total cost is much lower.
I hadn't thought of this. Imagine Google buying garages all over NYC and filling them with driverless Smart cars (or similar). A button is added to your phone's Google Maps app, so that in addition to "Get Directions" and "Navigate", you also have the option of "Pick Me Up".
So, you search for where you want to go on Google Maps and tap "Pick Me Up", with the option of telling it how many people are in your party. Within 5 minutes enough Smart cars for your party stop within 20 feet of you, even if you've already started walking, thanks to your phone's GPS. You get in the car, which has no steering wheel, and touch "Go" on the in-dash touch-screen to start the car driving to your destination.
I could get used to that.
Driverless is not just doing things the way we have been but without operating the car. It changes everything. The economics. The culture. etc.
Google on the other hand have driven half a million miles with their solution.
Object recognition from video is probably the future, it's how we work after all. The title though? Quite misleading.
edit: I don't mean to dismiss the student's accomplishments, but wanted to bring some perspective
To mitigate the issues in complex systems like planes and so on, the systems are built with multiple redundancies in place. That way even if something fails, the system remains stable and operative. So you don't need to have a "per trillionth" (and you are using the wrong word) quality system in order to deliver a very robust system.
Or Google might take him to work for them :)
It's that last 2 percent that's the most challenging.
And for the entire fleet of autonomous cars, the system does not need to be safe. It just needs to be better than humans driving. Realistically to gain acceptance they would need to be an order of magnitude better, which translates to roughly 1 fatal accident per year per 10^5 self driving cars. The current numbers for the US are 36000 fatalities and 250M cars.
I really hope this is true. I really want self-driving cars everywhere, now. Road Traffic accidents kill very many people and this is an example of Google using money and smarts to do good.
But looking at the way people deal with risk makes me wary.
People drive cars all the time even though driving is a bit risky. People don't really understand how good or bad their own driving is. There's a bunch of cognitive biases and rationalisations.
I hope people who are experts in communications are ready to dispel the FUD backlash against self driving cars.
One of the problems is insurance. Right now, everyone needs to have insurance to drive. If I accidentally hit a pedestrian or rear-end another driver they can claim on my insurance.
Let's say that a software bug results in someone getting hit and injured by a self-driving car. Who's liable? I am not sure that this is FUD, so much as a genuine question. It is because driving is risky that drivers must have insurance. Presumably most people would want self-driving cars to be insured if there was a chance of failure (even planes, one of the safest forms of transport, get insured against crashing).
We could very well have a situation where cars self drive, but drivers are still expected to pay attention to the road and 'drive' to avoid liability.
I suspect there's going to be a lot of lobbying on both sides before mass deployment of self driving cars.
On the other hand, I think that the politics of self driving cars could get rather interesting, especially if widespread adoption is fast, say 5 years until a self driving car is no longer extraordinary and 10 years until one just assumes that a new car is self driving. Then I think we get a initial phase, were they are seen as modern. After this, when every single accident gets widely reported, I think it is possible that the public completely splits, on one hand the traditionalists, who maintain that only a human should drive. And on the other hand the people who say that self driving cars are safer than humans. And the laws will shift according to the relative strength of the two groups. ( With probably some rather ridiculous political compromises in between, perhaps you are allowed to read a newspaper, but not to drink coffee. )
The same person who is liable if your current car has a fault tomorrow that causes you to crash, presumably.
"Student takes Google's self-driving car and offers the potential to make it $71,000 cheaper contingent upon correcting a misidentification issue that causes a possible failure rate of 6%"
Yeah, that would be so much better ;)
It's OK to take a little license with a title and try to summarize the main points. Maybe just adding the word "potential" in there would be enough, but some user named lololinternet would come along and claim how misleading the title is because ...
Definitely, congrats to the student! He really did astonishing work. But, nothing that could be useful in production. As I'm familiar with use of computer vision in traffic from academic and industrial point of view, I know situations this system has to deal with. And, I know what are the state-of-art results in that area. Computer vision is heavily used in traffic, but, self driving car is still out of its reach.
The Murray Gell-Mann Amnesia Effect
http://seekerblog.com/2006/01/31/the-murray-gell-mann-amnesi...
It'd also be interesting to hear about the difference between well funded laboratories (Google); Student labs; and commercial products.
It'd make an excellent post for HN if you ever have the time.
http://www.conti-online.com/www/automotive_de_en/themes/comm...
There's also a lot of work going on with driverless technology.
Disclaimer: I work for a Conti subsidiary that develops camera-based surround view and object detection systems but not directly on the products (IT Support)
http://www.asl360.co.uk/
Once you see examples from the datasets, lots of problems come to mind (and to todo list a bit later). This is a quite expensive problem to tackle with.
Some requirements: (1) you need datasets from various places, various weather conditions and various situations, (2) everything must work in realtime, (3) equipment is expensive (cameras, cars, gas, ...), (4) error has to be minimal (we are talking about human lifes). (and this is just part of requirements)
Student labs fail at money part, companies fail at lack of time (again, money; you need lots of time to deal with extreme number of situations and produce error prune product -- and nobody guarantees that you'll manage to do that).
And now, some problems: - Everything has to work in realtime which means more than 30 FPS in average (you need to aim for higher average speed so you don't get lags in complex scenes). Computer vision algorithms aren't usually realtime, for example, for simple task as object detection one of the best realtime algorithms is Viola-Jones which is more than 10 years old and patented. In newer days there are some breakthroughs in this area, but quite small if you take ten-year gap (and we are talking about quite simple problem -- object detection). - Then, the datasets. You need a lot of them. Different places, different times. All you can do is to beg someone for it, pay a lot for it or make an contract with traffic companies (for the product that you don't know will it work well; and good luck if you represent a student lab). - Now, the data. Take a ride during the different weather conditions and times of day or year. You'll be ok, but from CV point of view, you'll meet hundreds different problems. Disorted view during the rain, big balls of light during the night, different types of cars (cars with trailers, bikes at the back, motorbikes, ...), damaged road, damaged traffic control, traffic accidents, ... you get the point. - Then, think of number of things that you need to take care of -- traffic signs (very hard problem! specialy if you want to ride on local roads where some of signs are only partially visible), traffic fixes, etc.
Some of solutions are taking only some of the problems and tacking with them (more in way of alerting the driver). I'm not familiar with different sensors that deal with some of these problems, but maybe some of them aren't so expensive (today you can buy a car which can park itself (or that was just R&D showcase)).
Anyway, this is extremely interesting problem and it deserves us to fight with it. Unfortunately, from business side it looks like a big gamble.
Segfaults are going to have a really bad result.
In that context perhaps the problem is somewhat simpler if one assumes the infrastructure will be modified to help these cars. A lot could be done if each car communicated with roadside beacons and other locating systems. Some will voice concerns about privacy. I couldn't care less. Your movements are already far more traceable than they were fifty or a hundred years ago.
If I could manually get on the Golden State; engage auto pilot and be alerted a couple of miles before my exit in San Jose I'd be thrilled. I wouldn't even care if the thing pinged me every 15 minutes to see if I am awake.
Personally I would much rather have better above-ground transportation in the city than on the highway, but at that point I should be wishing for self-driving buses.
Props to this kid for being forward thinking, but this is a rather uninformed news article.
My question is, how far beneath the current human-driver accident rate will the AI-driver accident rate need to be before you accept that it is a better driver than you?
The article, like most articles uses an attention grabbing headline from the "$71,000 cheaper" point. The point of that is to give your brain a low resolution picture of the story. Enough to interest you and get you to read. Regardless of how you feel about that, if it worked (you are reading the article) it's best to forget about the conclusions/objections you started developing from the headline after you read the article. They're just a distraction.
Obviously, this isn't a drop in replacement for Google's system. Google probably haven't even been optimized their system for price anyway.
The interesting points here IMO is that (a)Google system cost ~$75k (b) a student was able to cheaply play in this space and get somewhere and (c) he did it by replacing the expensive 3d radar with a cheap 3d radar + webcam.
With so entrenched industry and distrusting public it will be uphill battle.
(not sure if serious)
What happens at night?