Says Tesla's approach is 360° flush-mounted tiny cameras plus radar, probably not lidar, with heavy image processing.
"Approach is 360 deg flush mounted tiny cameras + radar (prob not lidar)."
Why does Tesla use cameras instead of lidar?
Q13418 answers2013-2025
Short answer
He says Tesla bet on a pure AI solution using cameras and audio, while Waymo chose an expensive sensor suite.
All 18 answers, oldest to newest
He says once you solve passive optical, you have solved self-driving.
"once you solve passive optical, you've solved self driving"
He says SpaceX built its own lidar for Dragon docking, so he does not hate lidar, but roads require solving vision.
"So obviously if I hated lidar we would not have done that."
He says there is no question you can be superhuman with just cameras.
"There's no question that you can be superhuman with just cameras."
He says that for driving on real-world roads you have to solve vision.
"for driving on real world roads you have to solve vision."
He says eight-camera surround video will be superhuman because people do not have eyes in the back of their heads.
"because people don't have eyes at the back of their head."
Explains that vision plus radar trapped Tesla in a local maximum and pure vision matches how roads work.
"Going to pure vision set us back initially. Vision plus (coarse) radar had us trapped in a local maximum, like a level cap. Pure vision requires fairly advanced real-world AI, but that’s how our whole road system is designed to work: NN’s with vision."
Calls lidar a seductive local maximum; roads are designed for biological eyes, so self-driving needs cameras and neural nets.
"LIDAR is a seductive local maximum. SpaceX designed & built them to dock with ISS. However, the road system was designed to work with biological neural nets & eyes, so a general solution to self-driving necessarily will require silicon neural nets & cameras."
Says vision is hard but fundamentally superior to lidar.
"Vision does require solving real world AI, which is hard, but vision is fundamentally superior to lidar"
He argues the car can far exceed average human safety using vision alone.
"the car will, will far exceed the average human safety just with vision."
He explains Tesla turned off radar in Model 3 and Y because it generated more noise than signal.
"we turned it off because the radar actually generated more noise than signal."
He says the car looks in all directions at once and the computer never gets tired or distracted.
"the car is looking at all directions at once and we don't have eyes in the back of my head."
He explains Tesla turned off radar because of the disambiguation problem between vision and radar.
"Otherwise you get this disambiguation problem between vision and radar."
He says a low-cost inference computer and standard cameras are enough for self-driving, with no lidar, radar or ultrasonic sensors.
"it is obvious that our solution with a relatively low cost inference computer and standard cameras can achieve self driving. No lidars, no radars, no ultrasonics, nothing."
He contrasts Tesla's generalized solution with Waymo's localized, high-density-mapping approach.
"they have a very localized solution that requires high density mapping."
He says SpaceX built its own lidar to dock Dragon with the space station, so he has nothing against it, but it is not the right solution for cars.
"we literally designed and made a lidar to dock the space station. If I thought it was the right solution for cars, I would do that. But it isn't."
He says the digital equivalent of eyes and a brain are cameras and visual neural nets.
"the digital equivalent of eyes and a brain are cameras and, and, and visual neural nets or AI."
He says Tesla bet on a pure AI solution using cameras and audio, while Waymo chose an expensive sensor suite.
"we're the ones that made the bet that a pure AI solution with cameras and audio"