Tesla's AEB system fails according to Jalopnik/Luminar. Beware of FUD

[rob1987]

It seems like this thread could benefit from actually reading Tesla own description of what AEB does. It does not avoid collisions and is not designed to. It is designed to lessen the severity of collisions by slowing down. As such it only deploys once the collision is unavoidable i.e. stopping distance is less than distance to obstacle:

Getting hit by a car travelling 36 km/h is less severe than being hit by one at 52 km/h (roughly 15% fatallity rate at 52 km/h compared to roughly 2% at 36 km/h):

And yes, pressing the accelerator hard will override the AEB, as per the screen grab from the Tesla website above.

So the video is an accurate depiction of the AEB doing exactly what it was designed to do.

 

[AndreP]

2022 Tesla Model 3 4-door sedan

2022 Tesla Model 3 4-door sedan - midsize luxury car

www.iihs.org

2022 Tesla Model Y 4-door SUV

2022 Tesla Model Y 4-door SUV - midsize luxury SUV

www.iihs.org

BTW, "Superior" is their highest rating . . . .

Yup and lots of vehicles receive the "Superior" rating

Someone should go to CES, check this out, and report back with their own interpretation of any intentional or unintentional oversights, shortcomings, etc

 

[alexgr]

What does the manual say?

So, it is case #3. My guess is that the car went into an obstacle-aware acceleration mode when the driver kept pressing the accelerator pedal.

 

[AndreP]

We're speculating on the accelerator pedal status in a video posted months ago

The media right now is centered around a test track set up at CES you can visit at this very moment and experience for yourself in a Lexus equipped with this system versus Tesla, Audi, and whatever other systems they're exhibiting. If they're trying to pull the wool over people's eyes, it should be pretty apparent.

 

[alexgr]

We're speculating on the accelerator pedal status in a video posted months ago

The media right now is centered around a test track set up at CES you can visit at this very moment and experience for yourself in a Lexus equipped with this system versus Tesla, Audi, and whatever other systems they're exhibiting

At "this very moment" I am over a thousand miles from CES.

Looking at their Luminar-laced Lexus, you can see it stops at minimum one full car body from the target, which means the AEB has fully engaged deceleration at about 4 full car bodies at 30 mph. I doubt anyone wants the AEB work this way or we won't be able to drive on any roads in traffic.

 

[CyberGus]

Should the accelerator override AEB? If you can push through AEB and run something over regardless, doesn't that kinda defeat the purpose?

I refuse to drive a vehicle that overrides my judgement.

 

[AndreP]

I refuse to drive a vehicle that overrides my judgement.

Being able to override it through specific actions makes sense, sounds like you need to fully commit to the accelerator to override it in a Tesla? But I mean the AEB is designed to stop you from hitting something you don't see and where you'll likely have the accelerator depressed when that happens.

Any shortcomings of this should be relatively obvious to testers who work in the industry or follow cars / autonomous tech. I'd consider the video posted four months ago as shady if you weren't able to just go and try this out yourself at CES. They might still have rules at CES that make it easier to hide issues with the narrative they're trying to push, but that should stand out too.

 

[TresLA]

From my experience with the FSD beta 10.3 snafu, I was able to override the repeated false AEB with the accelerator. I don't think it was just FCW, but maybe someone else remembers better.

As for that video (from 4 months ago), it's really suspicious at 12-16 seconds where the alert sounds beep, then AFTERWARD the speed accelerates from 52 to 53 MPH momentarily before the car actually starts to brake. If not shady, perhaps it's just testing a system that wasn't designed to accomplish the same thing (i.e. - reduce injury in case of impact vs prevent impact). It's not just Tesla's AEB, but the industry standard to reduce the damage/injury of an imminent collision.

 

[CyberGus]

Being able to override it through specific actions makes sense, sounds like you need to fully commit to the accelerator to override it in a Tesla? But I mean the AEB is designed to stop you from hitting something you don't see and where you'll likely have the accelerator depressed when that happens.

Remember the 737-MAX? It had an automated stall-detection system that would point the aircraft downward to increase speed. Unfortunately, it's failures caused two planes to merrily slam into the ground while the pilots were frantically pulling back on the stick.

At a minimum, that is a violation of the 2nd law of robotics

 

[thesmokingman]

Too many contrarians who have no experience with these cars.

 

[Daniel in SD]

Remember the 737-MAX? It had an automated stall-detection system that would point the aircraft downward to increase speed. Unfortunately, it's failures caused two planes to merrily slam into the ground while the pilots were frantically pulling back on the stick.

At a minimum, that is a violation of the 2nd law of robotics

Are ABS and stability control also a violation of the 2nd law of robotics?
All these active safety systems override driver input...

 

[Daniel in SD]

Too many contrarians who have no experience with these cars.

True, a suspect no one posting here has attempted to run into something at 30mph in their Tesla.

 

[AndreP]

Remember the 737-MAX? It had an automated stall-detection system that would point the aircraft downward to increase speed. Unfortunately, it's failures caused two planes to merrily slam into the ground while the pilots were frantically pulling back on the stick.

At a minimum, that is a violation of the 2nd law of robotics

There are some obvious differences here in terms of risk from a system that sends you hurtling towards something versus a system that stops your motion (and doesn't depend on motion to keep you in their air), but slamming on the brakes in a vehicle comes with obvious risks as well.

I have a feeling Luminar is showing the capability of the technology here rather than a system that would actually be deployed on public roads. And production versions of similar systems in comparison vehicles would likely not be geared the same way because of all these nuances, so at best I'd guess it's not apples-to-apples.

 

[Daniel in SD]

There are some obvious differences here in terms of risk from a system that sends you hurtling towards something versus a system that stops your motion (and doesn't depend on motion to keep you in their air), but slamming on the brakes in a vehicle comes with obvious risks as well.

I have a feeling Luminar is showing the capability of the technology here rather than a system that would actually be deployed on public roads. And production versions of similar systems in comparison vehicles would likely not be geared the same way because of all these nuances, so at best I'd guess it's not apples-to-apples.

Exactly, I'm sure Tesla could tune the AEB system to work in these tests 100% of the time. Unfortunately such a system would have way more false positives (i.e. phantom braking). AEB systems are tuned to have a very low false positive rate, we have no idea what the false positive rate is for Luminar's system.

 

[AndreP]

Exactly, I'm sure Tesla could tune the AEB system to work in these tests 100% of the time. Unfortunately such a system would have way more false positives (i.e. phantom braking). AEB systems are tuned to have a very low false positive rate, we have no idea what the false positive rate is for Luminar's system.

That would definitely be my impression as well

 

[rob1987]

I just got through watching the joint Luminar-Volvo-Zenseact press conference from the CEC. This might be petty, but the first thing that struck me is how poorly a company specialized in the collection and interpretation of visual and spatial data was at mastering simple audio-visuals. They play different songs over the top of each other and some video issue causes all motion to be rendered as though it's viewed through a prism grating:

I certainly hope their AI isn't looking at a world rendered like that.

They don't give any specifics as to how often their AEB inappropriately deploys during driving. Since their demonstration scenarios seem contrived to show only that particular feature, I highly suspect their models have a bias towards harder braking in more situations, including inappropriate ones. A philosophy of "better to brake when we didn't need to than not brake when we needed to".

 

[CyberGus]

Are ABS and stability control also a violation of the 2nd law of robotics?
All these active safety systems override driver input...

Someone slamming the brakes is attempting to stop as quickly as possible, not lose control in a locked-up skid. ABS et al are designed to give the driver their desired result.

Presumptuous, perhaps, but within Asimov's laws

 

[Daniel in SD]

Someone slamming the brakes is attempting to stop as quickly as possible, not lose control in a locked-up skid. ABS et al are designed to give the driver their desired result.

Presumptuous, perhaps, but within Asimov's laws

That's why when you want to run over pedestrians you can turn off AEB.

 

[alexgr]

That's why when you want to run over pedestrians you can turn off AEB.

Those pesky pedestrians are destined to rid selves from the genetic pool of humanity.