10 ms·

Google Driverless Car – The Obstacle Detection Unit

63 points by todd8 12y ago

pkorzeniewski 12y ago
I wonder how much a production version will cost, with all the high-tech equipment..
- Symmetry 12y ago
  It looks like the lidar costs $75K retail, but none of the rest of the sensors look particularly expensive.
  - DannyBee 12y ago
    This is because they produce so few of them a year. If someone wanted to sit around and just produce lidar units, i'm sure it could be done for 100x less.
  - Rthakur 12y ago
    Velodyne recently released a 16 channel LIDAR for ~ $8000. I agree with you that the cost can come down tremendously with volume. You also need to consider the differences within LIDAR systems. The velodyne system is for 360 degree field of view. Other automotive LIDAR systems (See continental, Valeo, Ibeo, etc) with lower FOV and functionality (forward collision warning, blindspot detection,etc) are much lower cost > 10x ....
- lukasm 12y ago
  even 0.5m would make it stainable business.
- seren 12y ago
  For a business, e.g. a taxi, you'll have to compare it with a driver salary for X years.
Symmetry 12y ago
I wonder if the radars they're using can do frequency binning, that is figuring out the relative velocity of objects they detect via Doppler as well as their distance? Also, that 64 beam 360 degree lidar makes me drool a little bit.
- vonmoltke 12y ago
  I don't see why not. Pulsed transmitters are fairly ubiquitous nowadays. A bit more complicated and expensive than CW transmitters, but easy nonetheless.
dmm 12y ago
If lots of these cars were on the road would these sensors interfere with each other?
- TeMPOraL 12y ago
  Probably a bit, yes, but there are ways to filter out the interference to some extent. Anyway, I'm strongly hoping that they'll make cars that talk to each other - so instead many cars adding noise to each other's readings, we'll have cars that share data and dynamically delegate the task of detailed region mapping to few cars that would not interfere with each other.
  - aetherson 12y ago
    How would you handle malicious actors in such a situation?
    - meowface 12y ago
      That's a good question, and it makes me think that for that reason, Google probably won't actually be able to implement what the above post is suggesting. The most that can happen is each car will probably end up sending Google certain diagnostic information, but actual environmental sensing will have to be done by the car itself, with maybe some help from aerial cameras or satellites that Google controls.
    - jonlucc 12y ago
      Maybe by voting? It probably shouldn't be used for the most vital aspects, but it would be less likely to have 5 or 6 malicious cars around than just 1.
      - icebraining 12y ago
        Unless there are no real cars, just a device pretending to be many. Maybe we need a proof-of-car-presence :)
TeMPOraL 12y ago
Offtopic, but it kind of stood out for me in the text: > driverless car (i.e. not "self-driving car") > robotic chauffeur Is it me, or the article is pushing some new terminology?
- icebraining 12y ago
  Driverless is not new, e.g. http://en.wikipedia.org/wiki/Google_driverless_car http://en.wikipedia.org/wiki/Google_driverless_car
  - ape4 12y ago
    I prefer driverless horseless carriage.
    - jeremycole 12y ago
      So just "carriage" then? :)
sharemywin 12y ago
This is one of the places where I think there should be colaboration across companies. I'm not sure "not hitting pedestians" should be market differentiator.
- icebraining 12y ago
  It's not a "market differentiator". If a company can't develop that "feature", they just won't be able to release their driverless cars.
  - mrec 12y ago
    It's not a binary checkbox feature, though. I doubt any implementation will make it utterly impossible to hit a pedestrian - and that's fine, as long as they do better than meatbag drivers. A better detection/avoidance implementation is absolutely a differentiator, mostly via lower insurance premiums for operators.
    - JoeAltmaier 12y ago
      That's not clear, in our litigenous society. I hit a pedestrian: I get in trouble. A driverless car hits a pedestrian: all driverless cars (of that manufacturer) are potentially 'at fault' and recalled. Its definitely a different bar to pass.
      - mrec 12y ago
        But recalls happen already, for all sorts of reasons. They're expensive, and they can have a big impact on a manufacturer's reputation, but they don't mean that nobody can sell a car in the first place.
      - josho 12y ago
        That's interesting. A computer hits a pedestrian, then presumably the algorithms are updated and the probability of that scenario playing out again drops to 0. While A human hits a pedestrian, and that scenario is likely to never occur for that driver again, but for all other drivers the probability remains the same.
r3m6 12y ago
Is Lidar really required? I assume Google uses it only because it does not care about cost at this point of the development. German car makes (Mercedes, Audi, BMW,...) think they can create a driverless car without the need for expensive Lidar, see for example http://spectrum.ieee.org/cars-that-think/transportation/self-driving/bertha-the-car-comes-to-california http://spectrum.ieee.org/cars-that-think/transportation/self...
- cdash 12y ago
  Would Lidar really be all that expensive if it was used in mass produced driverless cars?
  - morley 12y ago
    Would anyone mass-produce a driverless car if Lidar were so expensive?
    - randyrand 12y ago
      Your questions both have the same answer (ie it's the same question) but that's a backwards way of thinking about it. You have to consider the cost of lidar at economics of scale, not before. If your point is that lidar won't be cheap enough even at economies of scale to make a mass market device, then just answer OPs question directly. Your comment is just a restatement in a less intuitive way.
  - ddeck 12y ago
    I'd be interested to hear from any experts in the are on this. Why is Lidar so expensive? My vacuum cleaner has Lidar. Clearly not of the same grade as that on the Google car, but what is it that makes their unit cost tends of thousands of dollars and is this likely to fall dramatically at scale?
    - benjamincburns 12y ago
      I don't have much direct experience with lidar, but with my background in in situ sensing I could probably land a job working with/on it. If I may be so bold as to presume that I just barely qualify as an "expert" here: The lidar in your vacuum (presumably a neato-XV series; if so, you have fine taste sir/madam) is no more than a spinning laser range finder. These work by offsetting a laser from a linear image sensor (imagine a camera with a crazy wide aspect ratio, maybe 256x6 resolution) and calculating distance by simple trigonometry or a pre-calibrated look-up table. These are called spinning parallax lidar. Their output is two dimensional in polar coordinates (angle, distance). More expensive lidar systems use carrier wave phase measurements (modulate the laser to a known signal, look for its reflection and calculate phase difference), direct time of flight measurements (like narrow beam sonic range finding, but with light, requires very expensive timing circuitry), doppler shift, and/or various structured light transformation/interference effects. Some units only use one of these techniques, some use a combination of these techniques. Most of these units operate using a number of lasers simultaneously to spit out 3 dimensional data, either as planar depth maps, or as 3D polar coordinates (angle-x, angle-y, radius). Either way, all of these are expensive because they require very fast and very precise complex analog circuitry, tight controls on noise sources (RF interference, temperature dependence, very good power isolation), tight controls on spurious emissions (GHz oscillators), careful calibration, and most importantly, very complicated software. To me, time of flight is the most interesting, but the least practical/likely to be made inexpensive. For a DIY example of the kind of timing circuitry involved, check out [1]. That said, there's room to make it cheaper. One such way may be to use a structured light transformation technique similar to Kinect, but speed up the camera a bit and modulate the output to a known signal (gives better interference rejection). 1: http://piotr.nikiel.info/?page_id=95 http://piotr.nikiel.info/?page_id=95
Gravityloss 12y ago
The sensor fusion reminds me of the F-35 JSF. The software seems to be one of the big things in the fighter project. So actually this might be a good fit for Google after all. Sensors are getting trivialized. I'm sure Google is working for a LIDARless car. They've published their stereo vision to 3d library. https://code.google.com/p/openvis3d/ https://code.google.com/p/openvis3d/
- vonmoltke 12y ago
  Indeed. This is a sensor fusion problem I really want to work on. I loved the problem space while I was working for Raytheon, but I can't stand working on large government programs. I'm surprised about the lack of IR sensors, though. I can think of cases where the IR data could be fused with data from other sensors, particularly visual, to improve object recognition. Particularly pedestrian and animal detection.
  - toomuchtodo 12y ago
    > Indeed. This is a sensor fusion problem I really want to work on. Have you thought about working for Tesla? autopilot@teslamotors.com
    - vonmoltke 12y ago
      I have, actually. I'm concerned about the work-life balance there, though.
      - toomuchtodo 12y ago
        Same reason I don't work in IT at SpaceX. I wouldn't have a wife anymore.
- krasin 12y ago
  I should point out that openvis3d is relatively old news (2006-2007).
robbrown451 12y ago
I understand how the cars can be aware of other objects on the road, but are they actually aware of themselves and where they fit into the environment? It seems they'd have to be in order to function, right? So we've get self-aware machines now. Did anyone even notice?
- timdierks 12y ago
  By this calculus, a line-following robot [1] is self-aware, with a more limited "environment". [1]: http://www.instructables.com/id/Line-following-Robot-with-Arduino/ http://www.instructables.com/id/Line-following-Robot-with-Ar...
- Gravityloss 12y ago
  If it sees itself in a mirror and notices a stain, does it drive up to a shrub and start going back and forth to rub itself clean?
  - Rangi42 12y ago
    If it were programmed to keep itself clean, and had an algorithm to do that ("scrape the dirty area against a shrub"), I see no reason why it couldn't pass the mirror test.
nitin_flanker 12y ago
The car is really amazing. This blog has posted 5-6 posts about the driverless car. here is the link to the tag - http://greybmusings.wordpress.com/tag/driverless-car/ http://greybmusings.wordpress.com/tag/driverless-car/ Google is utilizing multiple techs in this invention.

63 points by todd8 12y ago

pkorzeniewski 12y ago

I wonder how much a production version will cost, with all the high-tech equipment..

Symmetry 12y ago
It looks like the lidar costs $75K retail, but none of the rest of the sensors look particularly expensive.
- DannyBee 12y ago
  This is because they produce so few of them a year. If someone wanted to sit around and just produce lidar units, i'm sure it could be done for 100x less.
- Rthakur 12y ago
  Velodyne recently released a 16 channel LIDAR for ~ $8000. I agree with you that the cost can come down tremendously with volume. You also need to consider the differences within LIDAR systems. The velodyne system is for 360 degree field of view. Other automotive LIDAR systems (See continental, Valeo, Ibeo, etc) with lower FOV and functionality (forward collision warning, blindspot detection,etc) are much lower cost > 10x ....
lukasm 12y ago
even 0.5m would make it stainable business.
seren 12y ago
For a business, e.g. a taxi, you'll have to compare it with a driver salary for X years.

Symmetry 12y ago

I wonder if the radars they're using can do frequency binning, that is figuring out the relative velocity of objects they detect via Doppler as well as their distance? Also, that 64 beam 360 degree lidar makes me drool a little bit.

vonmoltke 12y ago
I don't see why not. Pulsed transmitters are fairly ubiquitous nowadays. A bit more complicated and expensive than CW transmitters, but easy nonetheless.

dmm 12y ago

If lots of these cars were on the road would these sensors interfere with each other?

TeMPOraL 12y ago
Probably a bit, yes, but there are ways to filter out the interference to some extent. Anyway, I'm strongly hoping that they'll make cars that talk to each other - so instead many cars adding noise to each other's readings, we'll have cars that share data and dynamically delegate the task of detailed region mapping to few cars that would not interfere with each other.
- aetherson 12y ago
  How would you handle malicious actors in such a situation?
  - meowface 12y ago
    That's a good question, and it makes me think that for that reason, Google probably won't actually be able to implement what the above post is suggesting. The most that can happen is each car will probably end up sending Google certain diagnostic information, but actual environmental sensing will have to be done by the car itself, with maybe some help from aerial cameras or satellites that Google controls.
  - jonlucc 12y ago
    Maybe by voting? It probably shouldn't be used for the most vital aspects, but it would be less likely to have 5 or 6 malicious cars around than just 1.
    - icebraining 12y ago
      Unless there are no real cars, just a device pretending to be many. Maybe we need a proof-of-car-presence :)

TeMPOraL 12y ago

Offtopic, but it kind of stood out for me in the text: > driverless car (i.e. not "self-driving car") > robotic chauffeur Is it me, or the article is pushing some new terminology?

icebraining 12y ago
Driverless is not new, e.g. http://en.wikipedia.org/wiki/Google_driverless_car http://en.wikipedia.org/wiki/Google_driverless_car
- ape4 12y ago
  I prefer driverless horseless carriage.
  - jeremycole 12y ago
    So just "carriage" then? :)

sharemywin 12y ago

This is one of the places where I think there should be colaboration across companies. I'm not sure "not hitting pedestians" should be market differentiator.

icebraining 12y ago
It's not a "market differentiator". If a company can't develop that "feature", they just won't be able to release their driverless cars.
- mrec 12y ago
  It's not a binary checkbox feature, though. I doubt any implementation will make it utterly impossible to hit a pedestrian - and that's fine, as long as they do better than meatbag drivers. A better detection/avoidance implementation is absolutely a differentiator, mostly via lower insurance premiums for operators.
  - JoeAltmaier 12y ago
    That's not clear, in our litigenous society. I hit a pedestrian: I get in trouble. A driverless car hits a pedestrian: all driverless cars (of that manufacturer) are potentially 'at fault' and recalled. Its definitely a different bar to pass.
    - mrec 12y ago
      But recalls happen already, for all sorts of reasons. They're expensive, and they can have a big impact on a manufacturer's reputation, but they don't mean that nobody can sell a car in the first place.
    - josho 12y ago
      That's interesting. A computer hits a pedestrian, then presumably the algorithms are updated and the probability of that scenario playing out again drops to 0. While A human hits a pedestrian, and that scenario is likely to never occur for that driver again, but for all other drivers the probability remains the same.

r3m6 12y ago

Is Lidar really required? I assume Google uses it only because it does not care about cost at this point of the development. German car makes (Mercedes, Audi, BMW,...) think they can create a driverless car without the need for expensive Lidar, see for example http://spectrum.ieee.org/cars-that-think/transportation/self-driving/bertha-the-car-comes-to-california http://spectrum.ieee.org/cars-that-think/transportation/self...

cdash 12y ago
Would Lidar really be all that expensive if it was used in mass produced driverless cars?
- morley 12y ago
  Would anyone mass-produce a driverless car if Lidar were so expensive?
  - randyrand 12y ago
    Your questions both have the same answer (ie it's the same question) but that's a backwards way of thinking about it. You have to consider the cost of lidar at economics of scale, not before. If your point is that lidar won't be cheap enough even at economies of scale to make a mass market device, then just answer OPs question directly. Your comment is just a restatement in a less intuitive way.
- ddeck 12y ago
  I'd be interested to hear from any experts in the are on this. Why is Lidar so expensive? My vacuum cleaner has Lidar. Clearly not of the same grade as that on the Google car, but what is it that makes their unit cost tends of thousands of dollars and is this likely to fall dramatically at scale?
  - benjamincburns 12y ago
    I don't have much direct experience with lidar, but with my background in in situ sensing I could probably land a job working with/on it. If I may be so bold as to presume that I just barely qualify as an "expert" here: The lidar in your vacuum (presumably a neato-XV series; if so, you have fine taste sir/madam) is no more than a spinning laser range finder. These work by offsetting a laser from a linear image sensor (imagine a camera with a crazy wide aspect ratio, maybe 256x6 resolution) and calculating distance by simple trigonometry or a pre-calibrated look-up table. These are called spinning parallax lidar. Their output is two dimensional in polar coordinates (angle, distance). More expensive lidar systems use carrier wave phase measurements (modulate the laser to a known signal, look for its reflection and calculate phase difference), direct time of flight measurements (like narrow beam sonic range finding, but with light, requires very expensive timing circuitry), doppler shift, and/or various structured light transformation/interference effects. Some units only use one of these techniques, some use a combination of these techniques. Most of these units operate using a number of lasers simultaneously to spit out 3 dimensional data, either as planar depth maps, or as 3D polar coordinates (angle-x, angle-y, radius). Either way, all of these are expensive because they require very fast and very precise complex analog circuitry, tight controls on noise sources (RF interference, temperature dependence, very good power isolation), tight controls on spurious emissions (GHz oscillators), careful calibration, and most importantly, very complicated software. To me, time of flight is the most interesting, but the least practical/likely to be made inexpensive. For a DIY example of the kind of timing circuitry involved, check out [1]. That said, there's room to make it cheaper. One such way may be to use a structured light transformation technique similar to Kinect, but speed up the camera a bit and modulate the output to a known signal (gives better interference rejection). 1: http://piotr.nikiel.info/?page_id=95 http://piotr.nikiel.info/?page_id=95

Gravityloss 12y ago

The sensor fusion reminds me of the F-35 JSF. The software seems to be one of the big things in the fighter project. So actually this might be a good fit for Google after all. Sensors are getting trivialized. I'm sure Google is working for a LIDARless car. They've published their stereo vision to 3d library. https://code.google.com/p/openvis3d/ https://code.google.com/p/openvis3d/

vonmoltke 12y ago
Indeed. This is a sensor fusion problem I really want to work on. I loved the problem space while I was working for Raytheon, but I can't stand working on large government programs. I'm surprised about the lack of IR sensors, though. I can think of cases where the IR data could be fused with data from other sensors, particularly visual, to improve object recognition. Particularly pedestrian and animal detection.
- toomuchtodo 12y ago
  > Indeed. This is a sensor fusion problem I really want to work on. Have you thought about working for Tesla? autopilot@teslamotors.com
  - vonmoltke 12y ago
    I have, actually. I'm concerned about the work-life balance there, though.
    - toomuchtodo 12y ago
      Same reason I don't work in IT at SpaceX. I wouldn't have a wife anymore.
krasin 12y ago
I should point out that openvis3d is relatively old news (2006-2007).

robbrown451 12y ago

I understand how the cars can be aware of other objects on the road, but are they actually aware of themselves and where they fit into the environment? It seems they'd have to be in order to function, right? So we've get self-aware machines now. Did anyone even notice?

timdierks 12y ago
By this calculus, a line-following robot [1] is self-aware, with a more limited "environment". [1]: http://www.instructables.com/id/Line-following-Robot-with-Arduino/ http://www.instructables.com/id/Line-following-Robot-with-Ar...
Gravityloss 12y ago
If it sees itself in a mirror and notices a stain, does it drive up to a shrub and start going back and forth to rub itself clean?
- Rangi42 12y ago
  If it were programmed to keep itself clean, and had an algorithm to do that ("scrape the dirty area against a shrub"), I see no reason why it couldn't pass the mirror test.

nitin_flanker 12y ago

The car is really amazing. This blog has posted 5-6 posts about the driverless car. here is the link to the tag - http://greybmusings.wordpress.com/tag/driverless-car/ http://greybmusings.wordpress.com/tag/driverless-car/ Google is utilizing multiple techs in this invention.