You can see the CAS system kick in if you watch the video closely (nose of the car dips) about maybe 0.7 seconds before the collision.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla on Autopilot slams into stationary car (VIDEO)
- Thread starter 2101Guy
- Start date
ArchHamster
Member
Who is green even? It sounds like he salvages video footage from crashed teslas for some reason then reports them, so we don't even have a connection to the original driver telling us in his words what happened?
It's really annoying digging through his tweets trying to find info on this. He said the tesla was on 2022.4.5.3 so not fsdb. I thought I saw him say it wasn't on fsd either but I can't find the tweet now in his sea of salvaged video tweets. I could see old tacc screwing up like this, but since this isn't fsd related and we don't have the testimony from the original driver, I kind of lost interest. All of this is speculation until the driver comes forward, imo. Until then, we are missing critical pieces of the story. Like isn't it possible to completely disable emergency breaking?
It's really annoying digging through his tweets trying to find info on this. He said the tesla was on 2022.4.5.3 so not fsdb. I thought I saw him say it wasn't on fsd either but I can't find the tweet now in his sea of salvaged video tweets. I could see old tacc screwing up like this, but since this isn't fsd related and we don't have the testimony from the original driver, I kind of lost interest. All of this is speculation until the driver comes forward, imo. Until then, we are missing critical pieces of the story. Like isn't it possible to completely disable emergency breaking?
Knightshade
Well-Known Member
Green is an extremely well known hacker in the Tesla world and has, for years, provided significant insights into what Tesla is doing on the computer/AP/etc side of things--- he also reviews salvaged computers for interesting info such as this crash.
It was on a divided controlled access highway-- FSDb would be entirely irrelevant since all public versions released still use to old AP/EAP highway stack there.
Yup not FSD Beta software. This crash from March was well before 2022.20.9 when Tesla switched radar 3/Y vehicles to Tesla Vision in August: Radar getting turned off on Model 3/Y with 2022.20.9
Maybe Tesla even has statistics showing Tesla Vision not using radar is safer on highways to avoid these types of crashes to decide to deploy it to vehicles with radar.
I wouldn't be surprised if they did have those stats. The radars Tesla has used are fairly useless in these situations. Not enough lateral precision to identify if the radar return is from a vehicle/object in your lane or beside it, so it's still reliant on the camera to be able to see the obstruction.
It’s a bit more complicated though, because I remember green mentioning they transferred AEB (and some other functionality) to pure vision some time before the full transition. Pretty sure this under-the-hood transition was happening in mid-late 2021.
AlanSubie4Life
Efficiency Obsessed Member
They won’t be, other than old reports on the accident that people dig up. This happened a while ago.
As mentioned it would not be FSD-related anyway.
This was from a car with radar, using a version of software without official TeslaVision, but perhaps some elements of vision system used for emergency features (speculative, know nothing about that).
Anyway you’ll find that FSDb will crash into things too. It’s just a matter of time before @verygreen gets the video - but the pipeline takes a while so could be many months from now for something that has already happened. We’ll see. Maybe no such accident has yet happened. But just a matter of time.
FSDb is not magic! The goal is to build a car that does not crash, but that’s the goal. There’s a reason you are driving the car when using FSDb. It’s intended to help assist you, and perhaps clean up your mistakes (ideally - it’s not there yet and puts users in worse situations than they would normally be in, all the time). But it is helping you, not the other way around (since you’re driving).
DarkForest
10.69 made me a shill
If it was using radar, then that’s the end of the story.
Radar on any car is setup to throw out datapoints with a speed differential of more than around 40 MPH. This info is usually included in the user manual for a vehicle equipped with radar.
So if the driver was going 60 MPH as an example, the radar would have seen the stopped vehicle and thrown out the data as noise.
Radar on any car is setup to throw out datapoints with a speed differential of more than around 40 MPH. This info is usually included in the user manual for a vehicle equipped with radar.
So if the driver was going 60 MPH as an example, the radar would have seen the stopped vehicle and thrown out the data as noise.
There's an additional reason the RADAR would throw out the data: the vehicle was stopped. Doppler RADAR uses a frequency shift to get the speed of objects, and if the car is going 60 mph, everything that is stationary is going to be returning frequency shifts for objects coming toward the RADAR transmitter at 60 mph. And by everything I mean the road itself, the guard rails, the trees and buildings around the car, etc. So a stopped vehicle would be indistinguishable from all other stopped objects. If the RADAR is something other than a pure CW RADAR and is FMCW, it would be able to get distance measurements by modulating the frequency and looking at the time delay of the shift in the returned frequency but I'm not sure what the resolution of such a RADAR system is at a distance of more than say 50 meters.
And, speaking as an ex-RADAR techie and working EE..
It's not that the RADAR on the car is purely Doppler. It's not. Just so we're clear: A pure Doppler radar sends out a continuous signal. The received signal comes back, goes through a multiplexer, and gets mixed with the transmitted signal. When things are moving, the received signal is at a different frequency, proportional to the difference in speeds of the transmitter and receiver. Police speed RADARs work like that; they measure that difference in frequency, work in the speed of the police car, and can tell how fast the (potential) speeder is going.
So, what's going on with the older Tesla RADARs? It's sending out pulses and receiving those pulses. That tells the car how far away the targets are. Doing math on a particular target return over time can give closing or expanding distance. But there's a problem: Clutter.
The faster one is going, say, faster than 55 mph or something, the farther out one needs to see a car one is doing TACC upon. But the radar return goes down as the square of the distance; so, when one is traveling fast, the RADAR return pulses from cars that one must track are very weak. And the return from stationary objects, like the ground, bridges, signs, guard rails, and parked cars that are close are going to overwhelm the radar return from the car one is trying to track. All that stuff, except the car one is trying to track, is termed, "clutter".
So, there's a solution. Even if the Tesla RADAR isn't a pure Doppler RADAR, the received signal does get shifted in frequency. So, one mixes the received GHz signal with the local oscillator in the car's transmit side RADAR; the result of the mixing of two signals like that (call them A and B) is the original two signals; the addition of the two signals (if one is transmitting at X.XX GHz, the result would be roughly 2*X.XX GHz), and the difference, A-B, which will be at baseband. And will be in the audio range, less than 50 kHz or so.
At this point, one hauls out a bog-standard low-pass filter (LPF) and runs the received signal through it. It's rigged so if radar returns are coming back from items whose velocity is 40 mph faster or slower than one is moving, those returns get filtered out, boom! And the stationary clutter from bridge abutments, guardrails, signs, parked cars, and all that jazz disappears, leaving only those targets traveling within +-40 mph of the car's velocity.
Result:
And, yeah, this happened. Several times. And now the biggies:
Nowadays, Teslas have switched from RADAR to vision. And I'm happy to report that with vision, we (a) get complaints that the car doesn't keep as steady a rate of speed as it used to do on freeways and (b) people have stopped complaining about how Teslas wouldn't stop for stopped objects.
Now, while reading this thread, there seems to be some clear indications that the car that crashed was (a) on a limited access highway, which means high speeds and (b) was using RADAR. No question: This is a driver error.
Finally: This is @2101Guy who started this thread. Um. This person seems to have a history of showing up with negative posts about Tesla. Enough said.
It's not that the RADAR on the car is purely Doppler. It's not. Just so we're clear: A pure Doppler radar sends out a continuous signal. The received signal comes back, goes through a multiplexer, and gets mixed with the transmitted signal. When things are moving, the received signal is at a different frequency, proportional to the difference in speeds of the transmitter and receiver. Police speed RADARs work like that; they measure that difference in frequency, work in the speed of the police car, and can tell how fast the (potential) speeder is going.
So, what's going on with the older Tesla RADARs? It's sending out pulses and receiving those pulses. That tells the car how far away the targets are. Doing math on a particular target return over time can give closing or expanding distance. But there's a problem: Clutter.
The faster one is going, say, faster than 55 mph or something, the farther out one needs to see a car one is doing TACC upon. But the radar return goes down as the square of the distance; so, when one is traveling fast, the RADAR return pulses from cars that one must track are very weak. And the return from stationary objects, like the ground, bridges, signs, guard rails, and parked cars that are close are going to overwhelm the radar return from the car one is trying to track. All that stuff, except the car one is trying to track, is termed, "clutter".
So, there's a solution. Even if the Tesla RADAR isn't a pure Doppler RADAR, the received signal does get shifted in frequency. So, one mixes the received GHz signal with the local oscillator in the car's transmit side RADAR; the result of the mixing of two signals like that (call them A and B) is the original two signals; the addition of the two signals (if one is transmitting at X.XX GHz, the result would be roughly 2*X.XX GHz), and the difference, A-B, which will be at baseband. And will be in the audio range, less than 50 kHz or so.
At this point, one hauls out a bog-standard low-pass filter (LPF) and runs the received signal through it. It's rigged so if radar returns are coming back from items whose velocity is 40 mph faster or slower than one is moving, those returns get filtered out, boom! And the stationary clutter from bridge abutments, guardrails, signs, parked cars, and all that jazz disappears, leaving only those targets traveling within +-40 mph of the car's velocity.
Result:
- If one is traveling slower than 40 mph or so, the distance one must see ahead to safely do TACC is reduced; the target is close, and therefore the target's return is nice and big, so TACC works. Clutter isn't particularly attenuated, but one doesn't care.
- If one is traveling faster than 40 mph or so, clutter gets attenuated by the LPF, and targets farther out can be seen, so TACC works.
And, yeah, this happened. Several times. And now the biggies:
- Every single other manufacturer of cars with TACC at that time also had this problem. Mercedes-Benz. Volvo. Toyota. You name it: They all had that LPF.
- Every single manufacturer of cars with TACC at that time had WARNINGS ALL OVER THE USER MANUAL. And that specifically included Tesla, where one had to acknowledge this characteristic of the car when enabling TACC.
Nowadays, Teslas have switched from RADAR to vision. And I'm happy to report that with vision, we (a) get complaints that the car doesn't keep as steady a rate of speed as it used to do on freeways and (b) people have stopped complaining about how Teslas wouldn't stop for stopped objects.
Now, while reading this thread, there seems to be some clear indications that the car that crashed was (a) on a limited access highway, which means high speeds and (b) was using RADAR. No question: This is a driver error.
Finally: This is @2101Guy who started this thread. Um. This person seems to have a history of showing up with negative posts about Tesla. Enough said.
Last edited:
Knightshade
Well-Known Member
Absolutely outstanding post from @Tronguy there... and yeah everyone using radar cruise has this issue- widely reported, and I know I've had to point out other-makers examples many times over the years when folks had not been aware of it.
It was... right up until AEB kicked on--- which was 2 seconds before impact- which you observed in the video as hard braking.
It was... right up until AEB kicked on--- which was 2 seconds before impact- which you observed in the video as hard braking.
AlanSubie4Life
Efficiency Obsessed Member
Most definitely. But over and over again people come out and are incredulous that a Tesla would do this, that it couldn’t have been on AP, etc. So it seems sort of like a valuable public service from a curmudgeon.
This is just a standard driver inattention accident. The driver was not paying attention and they, 100% driving the car (with AP convenience feature engaged until 2 seconds prior) slammed into a stationary object on the freeway. Happens every day. When 100% driving you should pay 100% attention!
Just to be clear: Tesla Vision equipped cars likely also do not stop for stopped objects in some cases. The most likely scenarios are ones where objects are difficult to identify due to lighting or other tricky special cases.
FSDb will still run into things. I don’t think we’ve seen a definitive serious case yet, but I’m quite confident that eventually we will. How often will this occur? No idea. Not sure what safety level is possible (if all the other glaring issues could somehow be solved and the system was limited by the sensors and object recognition).
But if you value your life (and the lives of others), you just drive the car yourself, as FSDb assists. As it is meant to be.
Last edited:
I’m no expert but to me it seems like it’s up to the sensor fusion implementation to discard such points and not something that happens at the radar level. Here’s a video I remembered demonstrating some stationary object returns at ~80mph:
Knightshade
Well-Known Member
if so what?
The only actually available stats we have are that cars running Teslas SW have accidents less often than cars without it.
There's absolutely issues one can raise with the data as far as not quite being apples to oranges so the added safety might not be as large as the #s imply, but I'm unaware of any data showing it's NOT safer than just a human.
With respect to what you said, above, true enough. But just to quibble: Finding those, "experimental" situations where the M3's RADAR would go wonky with stopped traffic wasn't difficult. Once warned, twice shy, and, on NoA/TACC/LK I used to intervene once or thrice on any, say, 200 mile trip. At least in the North East, it sure isn't unusual to be flying along at 60+ and discover stopped traffic ahead.
Since switching to Vision.. Well, the SO and I took a trip from NJ to FL and back in November. And while FSDb may have had its issues, the highway NoA/TACC/LK worked pretty much flawlessly (excepting the confusion about those pay-for tollways built into I-95 in Virginia) .. and there was never a hint that the car wasn't seeing stopped traffic.
Watts_Up
Well-Known Member
I wonder if any other Driver Assistance (or whatever name is used) from other companies encounter similar problem with stationary objects?
In particular, for Geofences based systems, if a lane has an obstacle not reported (like some temporary road work), would the system avoid it?
Last edited:
Knightshade
Well-Known Member
For consumer vehicles using radar?
Literally all of them.
This was just explained in posts 33 and 34 for example.
Was it ever confirmed which specific year/model this was? If it were a newly built performance model Y that was delivered without Brembo performance brakes, but standard brakes with the plastic covers to simulate performance brakes (and with no notification to customers that this was done) that Tesla quietly started doing, wonder if that could’ve impacted stopping distance any.
Last edited:
Similar threads
