3 ms·

Having LiDAR in any car for self driving seems like a prohibitively expensive idea though. For instance: See how no company has released a commercial product wi

by leesec 8y ago

Having LiDAR in any car for self driving seems like a prohibitively expensive idea though. For instance: See how no company has released a commercial product with LiDAR. And I definitely think Computer Vision will get better before LiDAR gets cost competitive.

Robotbeat 8y ago
I agree. However, suppose the opposite: Suppose LiDAR gets cheap first. Like, I don't know, $1000 per unit and compact enough to not be a big burden on the rest of the vehicle. Tesla charges $5000 for the FSD add-on. In principle, they could easily afford to retrofit vehicles with the cheap LiDAR plus maintain any advantages the full computer vision system would add (such as identifying vehicle types to assist in predictive behavior modeling and avoidance strategies). So trying to go the full Computer Vision (with ancillary radar) route actually has pretty low opportunity cost for Tesla. Another thing: I wonder how much simply much-improved geolocation could help? That failure from last year of driving right into a (failed) collision attenuator (where an off-ramp split off) could've been avoided simply by having high confidence, half-meter-or-better geolocation (from GPS, cell, IMU, etc fusion) and good mapping. And with good mapping/geolocation combined with other sensors, the sensors could focus on identifying changes to the expected road condition, perhaps increasing their robustness. (and when networked, they could use sensing from other vehicles ahead of the current vehicle to assist in navigation as well)
- leesec 8y ago
  AFAIK all the big companies working on this problem include high-definition mapping as a major component. The problem is gathering that data without having an actual LiDAR. So some companies like Waymo will literally just have cars with LiDARs drive all around every street and map it for them. This is great but means they're limited by when they travelled the road and conditions at the time. Ideally, you have a mapping system that is constantly updating (every time your fleet drives by something new like construction on the road it auto-updates the map). Comma.ai is working on a vision+GPS Kalman filter solution along these lines up to 10 cm accuracy. I'd guess Tesla is as well. And then yes, after that the vision would be primarily for localization.
  - Robotbeat 8y ago
    Indeed, as recent Tesla patent(s) show: https://electrek.co/2018/12/09/tesla-patent-technology-accurate-gps-positioning-for-self-driving-vehicles/ https://electrek.co/2018/12/09/tesla-patent-technology-accur...
syntaxing 8y ago
Totally agree. The race is a reputable solid state LiDAR which Waymo [1] and Velodyne announced a new one of recently [1]. But still, those are around 5K. Though I am really excited for Bajara's prism based scanner [3]. [1] https://techcrunch.com/2019/03/06/waymo-to-start-selling-standalone-lidar-sensors/ https://techcrunch.com/2019/03/06/waymo-to-start-selling-sta... [2] https://velodynelidar.com/velarray.html https://velodynelidar.com/velarray.html [3] https://www.baraja.com/ https://www.baraja.com/

by leesec 8y ago

Robotbeat 8y ago

I agree. However, suppose the opposite: Suppose LiDAR gets cheap first. Like, I don't know, $1000 per unit and compact enough to not be a big burden on the rest of the vehicle. Tesla charges $5000 for the FSD add-on. In principle, they could easily afford to retrofit vehicles with the cheap LiDAR plus maintain any advantages the full computer vision system would add (such as identifying vehicle types to assist in predictive behavior modeling and avoidance strategies). So trying to go the full Computer Vision (with ancillary radar) route actually has pretty low opportunity cost for Tesla. Another thing: I wonder how much simply much-improved geolocation could help? That failure from last year of driving right into a (failed) collision attenuator (where an off-ramp split off) could've been avoided simply by having high confidence, half-meter-or-better geolocation (from GPS, cell, IMU, etc fusion) and good mapping. And with good mapping/geolocation combined with other sensors, the sensors could focus on identifying changes to the expected road condition, perhaps increasing their robustness. (and when networked, they could use sensing from other vehicles ahead of the current vehicle to assist in navigation as well)

leesec 8y ago
AFAIK all the big companies working on this problem include high-definition mapping as a major component. The problem is gathering that data without having an actual LiDAR. So some companies like Waymo will literally just have cars with LiDARs drive all around every street and map it for them. This is great but means they're limited by when they travelled the road and conditions at the time. Ideally, you have a mapping system that is constantly updating (every time your fleet drives by something new like construction on the road it auto-updates the map). Comma.ai is working on a vision+GPS Kalman filter solution along these lines up to 10 cm accuracy. I'd guess Tesla is as well. And then yes, after that the vision would be primarily for localization.
- Robotbeat 8y ago
  Indeed, as recent Tesla patent(s) show: https://electrek.co/2018/12/09/tesla-patent-technology-accurate-gps-positioning-for-self-driving-vehicles/ https://electrek.co/2018/12/09/tesla-patent-technology-accur...

syntaxing 8y ago

Totally agree. The race is a reputable solid state LiDAR which Waymo [1] and Velodyne announced a new one of recently [1]. But still, those are around 5K. Though I am really excited for Bajara's prism based scanner [3]. [1] https://techcrunch.com/2019/03/06/waymo-to-start-selling-standalone-lidar-sensors/ https://techcrunch.com/2019/03/06/waymo-to-start-selling-sta... [2] https://velodynelidar.com/velarray.html https://velodynelidar.com/velarray.html [3] https://www.baraja.com/ https://www.baraja.com/