I heard from a tesla engineer that they do use lidar to train their models, but that they don’t see the value in deploying it to the fleet.
It’s worth watching the tesla event from about a year ago if you haven’t seen it for their full position: https://youtu.be/Ucp0TTmvqOE
The quick notes are that our roads are set up for human vision and a truly capable level 5 system will require human level interpretation of visual data (at which point lidar becomes redundant).
He thinks lidar with non human level visual interpretation won’t get us to full autonomy.
They see lidar as a short term half measure, a local maximum that’s a distraction from what has to get done anyway.
Humans also have only two eyes, so lets limit the AV to two visual spectrum only cameras, and while we're at it, use mainly the mirrors to look backwards (no repositioning cameras so only one looks backwards! Humans can't do that).
That may be enough for Level 5, but the entire idea is to exceed human driving capabilities (say, in dense fog). More sensors are an obvious advantage to use. Obviously not enough by themselves, but not a distraction either.
It's not intentionally limiting to match humans - it's recognizing that you need extremely high quality interpretation of visual data in order to achieve level 5 and if you have that then you don't need lidar because it has become redundant.
If you believe this argument to be true, then investment in lidar is a waste of time at best and at worst it's a distraction from getting the visual interpretation where it needs to be. Arguably companies that focus on it could be 'doomed' because of the investment cost/wrong focus, I think that's what they were getting at.
The dense fog is a good counter example - does lidar help in fog? I'd wonder if it would reflect off the fog?
Once you've got the visual part working at human level then adding additional lidar probably doesn't hurt (assuming you can get the costs down).
I don't agree it's a distraction, it's just more data for the model. More data should be good! There's enough money in this field to invest in everything* , and it's probably easier to add this in advance than to retrofit this in (Surely the retrofitted result would be inferior compared to just having LIDAR from the beginning so that the model can properly account for it).
As for dense fog, I am sure that even if LIDAR doesn't work, there's some other kind of no-human-equivalent sensor which does work.
Resource allocation plays a part no matter how many resources you have, any company wasting resources is likely to lose to one who doesn't.
In the specific case of LIDAR, it also increases the price for the end customer significantly, for negligible ( perhaps reaching 0 at the limit ) gain.
More data for the model means a bigger model using more computation and storage, and likely exponentially longer training time.
More sensors might be an advantage but cost is always a factor. The money spent on lidar might be better used on other cameras, CPU/GPU, or even safety features unrelated to self-driving.
I don't think it would be expensive, the same economy of scale that affects batteries would affect LIDAR (or any other sensor). Every other company can afford it, most of them without Tesla's valuation.
Besides, I don't think that AV companies at the stage where they can optimize for cost. For now they just need to get this thing working, and more data is better. The first winner will get so much cash V2 could then be optimized for cost.
Most people don't remember, but accelerometers used to be very expensive kit if you wanted to equip a drone or robot with one. Why did they get cheap? It wasn't phones, it was automotive airbags. If LIDAR can be made to work effectively then its price will be driven to the floor very quickly simply due to volume of production.
I don't think Tesla is truly anti-lidar, per se. In fact, I think if it was a practical technology they would add it to their toolbox.
But it's not practical. Lidar is currently EXTREMELY expensive.
And visual is better. It sees everything in all directions, and 99% of the same depth information lidar provides can be determined by two images separated in time.
As to safety, putting a bunch of inexpensive cameras on a car people can afford is way better than doubling the price of a car and making it unaffordable, just for some theoretical benefit beyond visual.
(teslas already have radar, long-range sonar and 8 cameras)
12 comments
[ 2.9 ms ] story [ 30.6 ms ] threadIt’s worth watching the tesla event from about a year ago if you haven’t seen it for their full position: https://youtu.be/Ucp0TTmvqOE
The quick notes are that our roads are set up for human vision and a truly capable level 5 system will require human level interpretation of visual data (at which point lidar becomes redundant).
He thinks lidar with non human level visual interpretation won’t get us to full autonomy.
They see lidar as a short term half measure, a local maximum that’s a distraction from what has to get done anyway.
That may be enough for Level 5, but the entire idea is to exceed human driving capabilities (say, in dense fog). More sensors are an obvious advantage to use. Obviously not enough by themselves, but not a distraction either.
It's not intentionally limiting to match humans - it's recognizing that you need extremely high quality interpretation of visual data in order to achieve level 5 and if you have that then you don't need lidar because it has become redundant.
If you believe this argument to be true, then investment in lidar is a waste of time at best and at worst it's a distraction from getting the visual interpretation where it needs to be. Arguably companies that focus on it could be 'doomed' because of the investment cost/wrong focus, I think that's what they were getting at.
The dense fog is a good counter example - does lidar help in fog? I'd wonder if it would reflect off the fog?
Once you've got the visual part working at human level then adding additional lidar probably doesn't hurt (assuming you can get the costs down).
I don't agree it's a distraction, it's just more data for the model. More data should be good! There's enough money in this field to invest in everything* , and it's probably easier to add this in advance than to retrofit this in (Surely the retrofitted result would be inferior compared to just having LIDAR from the beginning so that the model can properly account for it).
As for dense fog, I am sure that even if LIDAR doesn't work, there's some other kind of no-human-equivalent sensor which does work.
[EDIT: A LIDAR-based system to see through fog: https://www.asme.org/topics-resources/content/system-helps-s... ]
* ok, not everything, but most companies seem to do fine with LIDAR.
In the specific case of LIDAR, it also increases the price for the end customer significantly, for negligible ( perhaps reaching 0 at the limit ) gain.
More data for the model means a bigger model using more computation and storage, and likely exponentially longer training time.
So in that way they probably agree with you more than you think. I’d guess they find that to be the right trade off.
Besides, I don't think that AV companies at the stage where they can optimize for cost. For now they just need to get this thing working, and more data is better. The first winner will get so much cash V2 could then be optimized for cost.
But it's not practical. Lidar is currently EXTREMELY expensive.
And visual is better. It sees everything in all directions, and 99% of the same depth information lidar provides can be determined by two images separated in time.
As to safety, putting a bunch of inexpensive cameras on a car people can afford is way better than doubling the price of a car and making it unaffordable, just for some theoretical benefit beyond visual.
(teslas already have radar, long-range sonar and 8 cameras)