5 ms·

How about instead of using lidar, we use a camera on a swivel that can check mirrors and look out onto the road with basically the same visual sight as a human.

How about instead of using lidar, we use a camera on a swivel that can check mirrors and look out onto the road with basically the same visual sight as a human. Then for test data we just have our drivers wear some kind of camera glasses to capture what they are looking at and have them drive around, create labelled data in terms of what they see and how they respond in terms of input to the car. Then we can train a system that just kind of drives like humans do, with the same flaws, but hopefully with the same strengths as well.

wcoenen 7y ago
The advantage of lidar is not swivelling, but that it directly provides distance information (whereas with cameras you need to extract that from the parallax of two images). It's not clear to me what advantage a swivelling camera would bring; Tesla just has cameras for all directions, which is likely cheaper and more reliable.
- darepublic 7y ago
  hm yes I guess I misspoke when I said swivelling camera. But basically I was thinking of providing the same sorts of visual inputs as a human receives, and then training on actual human drivers driving.
freehunter 7y ago
Because people aren't going to trust self-driving cars unless they're drastically better than humans are. Being as good as a human will always lead to an inevitable crash where the human will say "I could have avoided that" and now the whole system crumbles. Human pedestrians get hit and killed by human-driven vehicles every day. Uber's self driving vehicle did it once, in a situation that was also missed by the human supervisor, and it was national news for weeks.
- cr0sh 7y ago
  > Because people aren't going to trust self-driving cars unless they're drastically better than humans are. I believe that people aren't ever going to trust self-driving cars, unless they are "perfect". Even ten 9s of reliability (no driving errors or accidents) would probably be too imperfect for people to accept (and I don't believe humans have ever made anything that close to perfection). I believe the reason this is probably so has to do with assignment of blame and guilt. Something easily done with human drivers, less so (or maybe impossible) for self-driving vehicles.
  - freehunter 7y ago
    Eventually I think they will, as the generations who drove their own car age out and new generations who grew up reliant on self-driving technology comes of age. You see it with automatic transmissions, some people don't trust the computer to shift gears for them because they think they can do it better. As time went on and automatic gearboxes became more reliable and more popular, those complaints slowed down. (speaking from an American standpoint here) You see it with anti-lock brakes (ABS). My grandparents still don't trust ABS even with studies showing ABS stops quicker than humans pumping the brakes, because it takes a level of control away from them. But that's not a common complaint these days. You see it with electric cars, where any fire make national news but no one seems to care about the thousands of ICE cars that burn to the ground every year. You even see it with seatbelts, people who claim it's safer to be thrown from the car in an accident. But eventually those complaints lose their novelty and people just accept it as normal.
- darepublic 7y ago
  Yes you may be right about what it will take to win public trust. But if I'm not mistaken self-driving cars haven't yet reached the ability of human drivers. If we were to just deploy them willy-nilly the incidents of bad-driving/accidents they cause would go up, no?
rjdagost 7y ago
There's a reason why almost everyone pursuing autonomous driving uses lidar despite the high costs. It's because lidar has extremely good "recall" and "precision" in most driving conditions. In other words, if an object is present on or near the road, lidar will almost certainly detect it (high recall), and it has very rare false positive detections (high precision). Lidar also directly provides distance information. For pure camera based approaches, we've seen huge improvements in recall in the past decade, but unfortunately we still need orders of magnitude reductions in the false negative rate to be good enough for safety critical applications. I think pure computer vision approaches will get there someday, but we first need some fundamental changes in CV algorithms in my opinion. Hopefully with increasing adoption, lidar will get significantly cheaper.

wcoenen 7y ago

The advantage of lidar is not swivelling, but that it directly provides distance information (whereas with cameras you need to extract that from the parallax of two images). It's not clear to me what advantage a swivelling camera would bring; Tesla just has cameras for all directions, which is likely cheaper and more reliable.

darepublic 7y ago
hm yes I guess I misspoke when I said swivelling camera. But basically I was thinking of providing the same sorts of visual inputs as a human receives, and then training on actual human drivers driving.

freehunter 7y ago

Because people aren't going to trust self-driving cars unless they're drastically better than humans are. Being as good as a human will always lead to an inevitable crash where the human will say "I could have avoided that" and now the whole system crumbles. Human pedestrians get hit and killed by human-driven vehicles every day. Uber's self driving vehicle did it once, in a situation that was also missed by the human supervisor, and it was national news for weeks.

cr0sh 7y ago
> Because people aren't going to trust self-driving cars unless they're drastically better than humans are. I believe that people aren't ever going to trust self-driving cars, unless they are "perfect". Even ten 9s of reliability (no driving errors or accidents) would probably be too imperfect for people to accept (and I don't believe humans have ever made anything that close to perfection). I believe the reason this is probably so has to do with assignment of blame and guilt. Something easily done with human drivers, less so (or maybe impossible) for self-driving vehicles.
- freehunter 7y ago
  Eventually I think they will, as the generations who drove their own car age out and new generations who grew up reliant on self-driving technology comes of age. You see it with automatic transmissions, some people don't trust the computer to shift gears for them because they think they can do it better. As time went on and automatic gearboxes became more reliable and more popular, those complaints slowed down. (speaking from an American standpoint here) You see it with anti-lock brakes (ABS). My grandparents still don't trust ABS even with studies showing ABS stops quicker than humans pumping the brakes, because it takes a level of control away from them. But that's not a common complaint these days. You see it with electric cars, where any fire make national news but no one seems to care about the thousands of ICE cars that burn to the ground every year. You even see it with seatbelts, people who claim it's safer to be thrown from the car in an accident. But eventually those complaints lose their novelty and people just accept it as normal.
darepublic 7y ago
Yes you may be right about what it will take to win public trust. But if I'm not mistaken self-driving cars haven't yet reached the ability of human drivers. If we were to just deploy them willy-nilly the incidents of bad-driving/accidents they cause would go up, no?

rjdagost 7y ago

There's a reason why almost everyone pursuing autonomous driving uses lidar despite the high costs. It's because lidar has extremely good "recall" and "precision" in most driving conditions. In other words, if an object is present on or near the road, lidar will almost certainly detect it (high recall), and it has very rare false positive detections (high precision). Lidar also directly provides distance information. For pure camera based approaches, we've seen huge improvements in recall in the past decade, but unfortunately we still need orders of magnitude reductions in the false negative rate to be good enough for safety critical applications. I think pure computer vision approaches will get there someday, but we first need some fundamental changes in CV algorithms in my opinion. Hopefully with increasing adoption, lidar will get significantly cheaper.