Real world Tesla camera based safety & self driving tech vs LiDAR

[ChrisO]

I mostly agree with you, but this is inaccurate. Many ASAD systems have (in)famously failed to recognize such things as "walls" whether they're other vehicles like tractor trailers or "actual" walls. People are accepting at face-value the "(full) self driving" claims and letting the vehicle take over when they shouldn't. It's a problem caused by both manufacturer and driver.

A system with "stereo vision" uses automated feature matching to determine the depth/distance to objects within the scene. Automated feature matching fails dramatically when it's presented with a flat, monotone, or otherwise featureless object. A brick wall that fills most of the field of view will look like nothing--like a blank scene; there will be no depth/distance information in that scene. Any 3D features that are in a wall are too small to match at the resolutions and speeds required for safety while driving.

Automated matching also fails when there are confusing signals such as flashing lights or repeated similar or identical objects. This is especially true when those flashing lights alternate colors or positions (safety vehicles). A flashing blue/red light in one scene that's either off or in a different place in the next scene can cause very weird distance/movement results even though the lights are not moving. Confusing results can also happen with a series of identical traffic control devices.

A poorly-designed safety system will not respond well to such scenes. An overly-permissive system might decide it's a glitch and continue driving. An overly-responsive system might slam on the brakes on a freeway when the vehicle is heading into the sun b/c the scene is washed out and the system decides it might be seeing a wall. A cautious system might yell at the driver to take over, but if they weren't paying attention, it may be too late. Figuring all of this out is not an easy task to program.

A system with two or more sensor types and a diversity of data would be able to cross reference the data and respond more appropriately.

Very good points. I was definitely letting the "fake wall" off too easy. The more I think about it, there are also the cases where a Tesla has run over a Harley Davison motorcycle because the back lights look like a car's back lights in the distance. That wouldn't happen with any system the can actually judge distance.

Sponsored

 

[SpaceEVDriver]

Very good points. I was definitely letting the "fake wall" off too easy. The more I think about it, there are also the cases where a Tesla has run over a Harley Davison motorcycle because the back lights look like a car's back lights in the distance. That wouldn't happen with any system the can actually judge "distance".

I was going to mention that as well, but my comment was already too long...

 

[Sikkun]

Logically auto pilot should hit the brakes if it is about to have a collision. I mean, how concerned are you about auto pilot function AFTER you crash?

Which is the question, did it disengage and hit the breaks? I believe it started stopping just not soon enough.

But if your in a minor collision and auto pilot stays on and the car just yolo keeps trying to drive…yeah that would be concerning.

 

[tobiasjef]

Tesla FSD Fails in Heavy Rain: Stormy Weather Test At Night a You tube video of FSD in heavy rain.

9 months ago. Tesla fixed these problems because they switched to end to end neural networks.

9 months old FSD was a several generations old version of FSD since was going through a major rework which they had been working on for almost 2 years.

 

[tobiasjef]

Very good points. I was definitely letting the "fake wall" off too easy. The more I think about it, there are also the cases where a Tesla has run over a Harley Davison motorcycle because the back lights look like a car's back lights in the distance. That wouldn't happen with any system the can actually judge distance.

no there hasn't. FSD has not hit a motorcycle or the motorcyclist would have been injured or dead.

We know about every death using FSD. there are exactly two of them. The first happened because a tractor trailor pulled out in front of the tesla and the tesla owners abused FSD by setting the speed to 70 in a 45.

If you mean "judging distance" you are talking about a mach e which radar cannot even see a stationary vehicle in broad daylight? One whose ADAS system is under investigation by the NHTSA for not seeing stationary objects?

 

[AhardFSU]

Look, Tesla is very irresponsible in the way that it promotes so called FSD. And you have a lot of YouTubers like Out Spec that will heap tons of praise on FSD and say that the car can drive itself, which is absolutely irresponsible and not true. But they’re so caught up in having fun and LOLs that they don’t talk responsibly. FSD has nowhere near the number of miles needed before a disconnect happens for it to be considered capable.

And Tesla trying to use cameras and “neural network” to pull off SAE level 3 self driving is irresponsible and more about cutting costs. It’s not about safety and people should stop promoting it as if it’s the next coming. I’m glad Mark released this video. A lot of folks need to be more circumspect in how they talk about Autopilot and FSD.

 

[tobiasjef]

Look, Tesla is very irresponsible in the way that it promotes so called FSD. And you have a lot of YouTubers like Out Spec that will heap tons of praise on FSD and say that the car can drive itself, which is absolutely irresponsible and not true. But they’re so caught up in having fun and LOLs that they don’t talk responsibly. FSD has nowhere near the number of miles needed before a disconnect happens for it to be considered capable.

And Tesla trying to use cameras and “neural network” to pull off SAE level 3 self driving is irresponsible and more about cutting costs. It’s not about safety and people should stop promoting it as if it’s the next coming. I’m glad Mark released this video. A lot of folks need to be more circumspect in how they talk about Autopilot and FSD.

FSD stops for the wall using HW4 and latest FSD build for cybertruck. Exactly like you would expect

FSD miles per intervention does not show the true progress of FSD. Driving improves while interventions happen for more trivial issues.

Tesla can literally solve most of their issues by fixing the entire map of the US. They wouldn't need to change a line of FSD code or anything with the neural network

 

[Sikkun]

I would suspect someone who joins a Mach E forum with an entire post history being about Tesla has probably drunk a lot of the kool aid.

 

[Billyk24]

FSD stops for the wall using HW4 and latest FSD build for cybertruck. Exactly like you would expect

FSD miles per intervention does not show the true progress of FSD. Driving improves while interventions happen for more trivial issues.

Tesla can literally solve most of their issues by fixing the entire map of the US. They wouldn't need to change a line of FSD code or anything with the neural network

FSD is yet to be dialed in. There is a reason why FSD has not obtained level #3 of autonomous driving in California and Neveda as Mercedes did. The previous reply about it was a bunch of bull and not related to the real reason which Mercedes states.

 

[ChrisO]

no there hasn't. FSD has not hit a motorcycle or the motorcyclist would have been injured or dead.

This is the first time I have seen where someone can predict that an accident didn't happen because there wasn't an injury or death. Believe it or not, the outcome doesn't have to be injury or death just because a motorcycle is involved.

But let's say this is "urban legend", I still condemn a company that uses the public highways for its beta testing, with untrained drivers at that.

I will say that I think the Telsa software is some of the most advanced in the world, but that doesn't give them the right to knowingly endanger the public.

 

[tobiasjef]

This is the first time I have seen where someone can predict that an accident didn't happen because there wasn't an injury or death. Believe it or not, the outcome doesn't have to be injury or death just because a motorcycle is involved.

But let's say this is "urban legend", I still condemn a company that uses the public highways for its beta testing, with untrained drivers at that.

I will say that I think the Telsa software is some of the most advanced in the world, but that doesn't give them the right to knowingly endanger the public.

Endanger the public?



Drive like this. I dare you. I bet you can't drive this well if you tried.

 

[tobiasjef]

But let's say this is "urban legend", I still condemn a company that uses the public highways for its beta testing, with untrained drivers at that.

Tesla uses monocular depth estimation. There are a lot of tricks but they are more advanced than judging a vehicle's size based on how far the tail lights are apart. Every car is a different size

 

[tobiasjef]

FSD is yet to be dialed in. There is a reason why FSD has not obtained level #3 of autonomous driving in California and Neveda as Mercedes did. The previous reply about it was a bunch of bull and not related to the real reason which Mercedes states.

Because you understand nothing about SAE driving levels or how to build a driving system. Your assumption (which is flat out idiotic) is that self driving cars go from level 2 to level 3 and then level 4.

Level 3 is not a useful form of autonomy which is why tesla has never aspired to ever release a level 3 product.

The SAE levels are not used by any self driving engineers because they are idiotic levels written by policy makers and not those who understand how the technology works.

No one uses the "level" of self driving as a metric for anything because it is pointless. Tesla has level 4 which shows how pointless the levels really are.

Funny how you repost 1 accident picture from over 4 billion miles of FSD driving. As if FSD is not getting into accidents as much as you would assume

 

[Jimrpa]

FSD stops for the wall using HW4 and latest FSD build for cybertruck. Exactly like you would expect

FSD miles per intervention does not show the true progress of FSD. Driving improves while interventions happen for more trivial issues.

Tesla can literally solve most of their issues by fixing the entire map of the US. They wouldn't need to change a line of FSD code or anything with the neural network

Sigh, does that mean we have to rename it to “the Gulf of Musk”? ???

 

[Billyk24]

Because you understand nothing about SAE driving levels or how to build a driving system. Your assumption (which is flat out idiotic) is that self driving cars go from level 2 to level 3 and then level 4.

Level 3 is not a useful form of autonomy which is why tesla has never aspired to ever release a level 3 product.

The SAE levels are not used by any self driving engineers because they are idiotic levels written by policy makers and not those who understand how the technology works.

No one uses the "level" of self driving as a metric for anything because it is pointless. Tesla has level 4 which shows how pointless the levels really are.

Funny how you repost 1 accident picture from over 4 billion miles of FSD driving. As if FSD is not getting into accidents as much as you would assume

It is amazing that you can not even realize why Tesla can not meet the California and Neveda requirements for level #3 autonomy.

Sponsored