11–20 of 222 posts
ITT: people who think they know better than the company that has been working on this stuff day in and day out for years lol.
Well, it's a great way to inflate the stock price. Much better than saying "We'll be just another car company, when self-driving arrives".
ITT: people who think they know better than the company that has been working on this stuff day in and day out for years lol.
Not having LiDAR for anything over level 3 self-driving capabilities seems like a very bad idea...Computer vision right now just does not have the spatial awareness that you need for self driving capabilities. I wish Tesla would work on something similar to the Kinect v3 but for long range (above 20 m).
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.
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)
Not having LiDAR for anything over level 3 self-driving capabilities seems like a very bad idea...Computer vision right now just does not have the spatial awareness that you need for self driving capabilities. I wish Tesla would work on something similar to the Kinect v3 but for long range (above 20 m).
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.
[1] https://techcrunch.com/2019/03/06/waymo-to-start-selling-sta... [2] https://velodynelidar.com/velarray.html [3] https://www.baraja.com/
Earlier quoted context omitted.
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.
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 pre…
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.
Earlier quoted context omitted.
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 pre…
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 mappi…
This is a (perhaps short-term) failure of first principles reasoning. Elon's known to favor thinking about things from first principles, and there's no theoretical reason that vision can't work. However, technical limitations might take a while to overcome.
ITT: people who think they know better than the company that has been working on this stuff day in and day out for years lol.
Or maybe they've gone down a blind alley in their rush to look like they were first to market.