I work with programs approaching L3+ from L2, with the requirement that the system works for 99% of roads (not tesla before people start fixating on that).
We find that the cases where lidar really helps are in gathering training data, parking, and if focused enough some long distance precision.
None of these have been instrumental in a final product; personally I suspect that many of the cars including lidar use it for data collection and edge cases more than as part of the driving perception model.
Waymo is the best current autonomous driving system and Waymo uses LIDAR. This is because LIDAR is an incredibly effective sensor for accurate range data. Vision and Radar range data is much less accurate and reliable.
Waymo used LIDAR in the realtime control loop. It combines LiDAR, camera, and radar data in real time to build a 3D representation of the environment, which is constantly updated.
I fundamentally don't trust any level 4 system that doesn't use LIDAR
Yes, I am aware of waymo... What they do is impressive. However they don't have a product that works for all highways yet, that's the space I work in, and we have no real fixation on lidar... It's nice but not a requirement, and hard to justify the cost unless you can make sales because of it (and there are some places where this is the case, but not everywhere)
You don't need the mm precision of lidar very often; we find that it offers nothing at speed over radar; and in tight manoeuvres the cameras we need for human park assist and ultrasonics do well enough.
It in not more accurate; but it is more precise, but that doesn't really matter. (Radar gives you relative speed directly, this is more important than a very precise point at highway speeds).
Sort of; accidents are the absolute core of the product. They are rare, but they are the focus of the design.
By edge cases I mean scenarios like the lights going out in an underground garage; low vision due to colourful smoke or dust, or things like optical illusions or occlusion that a human would just need to remember.
Lidar can help, but not really enough to be worth it.
Like Waymo? (https://dmnews.co.uk/waymo-robotaxi-spotted-unable-to-cross-...) 17 years after betting the farm on LIDAR the solution fails to navigate a puddle. Sorry but they bet on the wrong technology, Tesla has overtaken them with multi camera and NN solution.
Your conclusion from a single incident is a bad inference. One vehicle getting confused by a puddle (likely a sensor fusion edge case or mapping artifact, not a fundamental LIDAR failure) doesn't indict the technology. Tesla's cameras have produces vastly more failures.
Waymo has driven tens of millions of autonomous miles with a serious injury/fatality rate dramatically lower than human drivers. The actual data shows the technology works. Tesla FSD still requires active driver supervision and is not legally or technically a robotaxi system. Comparing them as if they're at parity is wrong.
LIDAR gives direct metric depth with no inference required. Camera-only systems must infer depth from 2D images using neural networks, which introduces failure modes LIDAR doesn't have. Radar is very valuable when LIDAR and cameras give ambiguous data.
What metrics has Telsa overtaken Waymo? Deployed robotaxi revenue miles? No. Disengagement rates? No published comparable data. Safety per mile in driverless operation? No.
A Tesla wouldn't stop for a puddle. Also its not locked to a small geofenced area (people have driven coast to coast without a single intervention on FSD including parking spot to parking spot) when I can buy a Waymo vehicle that does this then Waymo would have caught up with Tesla.
I work with programs approaching L3+ from L2, with the requirement that the system works for 99% of roads (not tesla before people start fixating on that).
We find that the cases where lidar really helps are in gathering training data, parking, and if focused enough some long distance precision.
None of these have been instrumental in a final product; personally I suspect that many of the cars including lidar use it for data collection and edge cases more than as part of the driving perception model.
Waymo is the best current autonomous driving system and Waymo uses LIDAR. This is because LIDAR is an incredibly effective sensor for accurate range data. Vision and Radar range data is much less accurate and reliable.
Waymo used LIDAR in the realtime control loop. It combines LiDAR, camera, and radar data in real time to build a 3D representation of the environment, which is constantly updated.
I fundamentally don't trust any level 4 system that doesn't use LIDAR
Yes, I am aware of waymo... What they do is impressive. However they don't have a product that works for all highways yet, that's the space I work in, and we have no real fixation on lidar... It's nice but not a requirement, and hard to justify the cost unless you can make sales because of it (and there are some places where this is the case, but not everywhere)
You don't need the mm precision of lidar very often; we find that it offers nothing at speed over radar; and in tight manoeuvres the cameras we need for human park assist and ultrasonics do well enough.
It in not more accurate; but it is more precise, but that doesn't really matter. (Radar gives you relative speed directly, this is more important than a very precise point at highway speeds).
1 reply →
Accidents are not normal driving situations but edge cases.
Sort of; accidents are the absolute core of the product. They are rare, but they are the focus of the design.
By edge cases I mean scenarios like the lights going out in an underground garage; low vision due to colourful smoke or dust, or things like optical illusions or occlusion that a human would just need to remember.
Lidar can help, but not really enough to be worth it.
4 replies →
Like Waymo? (https://dmnews.co.uk/waymo-robotaxi-spotted-unable-to-cross-...) 17 years after betting the farm on LIDAR the solution fails to navigate a puddle. Sorry but they bet on the wrong technology, Tesla has overtaken them with multi camera and NN solution.
Your conclusion from a single incident is a bad inference. One vehicle getting confused by a puddle (likely a sensor fusion edge case or mapping artifact, not a fundamental LIDAR failure) doesn't indict the technology. Tesla's cameras have produces vastly more failures.
Waymo has driven tens of millions of autonomous miles with a serious injury/fatality rate dramatically lower than human drivers. The actual data shows the technology works. Tesla FSD still requires active driver supervision and is not legally or technically a robotaxi system. Comparing them as if they're at parity is wrong.
LIDAR gives direct metric depth with no inference required. Camera-only systems must infer depth from 2D images using neural networks, which introduces failure modes LIDAR doesn't have. Radar is very valuable when LIDAR and cameras give ambiguous data.
What metrics has Telsa overtaken Waymo? Deployed robotaxi revenue miles? No. Disengagement rates? No published comparable data. Safety per mile in driverless operation? No.
A Tesla wouldn't stop for a puddle. Also its not locked to a small geofenced area (people have driven coast to coast without a single intervention on FSD including parking spot to parking spot) when I can buy a Waymo vehicle that does this then Waymo would have caught up with Tesla.
7 replies →
> Tesla has overtaken them with multi camera and NN solution.
Let me guess, you heard this from Elon?