Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Ross Gerber shows Dan O'Dowd that FSD is much safer than a human

This site may earn commission on affiliate links.
This is the moment Ross disengages FSD at time code 0:37.

How can that be? You just said he slammed on the brakes at the 36 second mark:

But watch the video again. At the 36 second mark, Ross slams on the brakes because

Make your mind up. Did he press the brakes at 36 or 37?

You can hear the chime when FSD beta disengages.
You are using a trailing indication which is about 2 seconds after disengagement.

You can see the blue path and the blue wheel on the screen.
No, you can see the path go from blue to grey at the 36 second mark.

At 35 seconds it is clearly blue:
1687570455950.png


at 36 seconds it has clearly changed to grey, and the animation has started to show the disengagement:
1687570513231.png


Here later, still at the 36 second mark, the path has gone to the thin grey line, while the steering wheel indicator has still not changed from blue because that animation lags the actual disengagement:
1687570617488.png

And he is clearly braking as the speed has dropped from 36 MPH to 31 MPH, before you say he disengaged.
 
Last edited:
FSDb is also supposed to not hit other cars. Stop sign or not it clearly would have plowed into the white SUV. You can even argue that the white SUV took the corner rather too sharply, I don't know or care if it did cut the corner. FSDb cannot hit other cars and it would definitely have hit that white car which had full right of way to be in the intersection.

The car comes to a much more drastic stop when Ross hit the brakes than a typical FSDb stop sign attempt. Thus the car was not going to stop. Odd that it said it was going to stop. But we've seen that in other videos where it shows the stop sign on the display and blows through it anyway.
 
I wonder if there is a bug in FSDb related to changing the set speed as it approaches a stop sign. Ross was running at 5 over the speed limit, but at second 33 he rolled it down to the speed limit, 35 MPH, right about the time the stop sign became visible. It still hadn't dropped to the new set speed by the time Ross started pressing the brakes.

I'll have to give that a try on my next drive to see if it impacts the stopping behavior.
 
Try watching the video in slow mo. You can see the path go from blue to grey, before the stop sign. The steering wheel is just in its flashing stage indicating the disengagement. Ross disengaged before the stop sign.

You can even see FSD saying it is going to stop for the stop sign before Ross disengages:

View attachment 950302

1687573425602.png

Yup...FSDb does say "Stopping for Stop Sign" but Ross took control right at that instant. Chimes are async and many ms delayed and cannot be taken as a control point.
 
The car is obviously late in planning the stop and, despite throwing up the message, would likely not have stopped in time. You can see the path of destruction extending through the intersection with no gray to indicate where the car plans to stop. In fact, the second screenshot shows a wide completely gray path of destruction. This is at the point of disengagement, but likely indicates, late, that the car has planned to stop at the stop sign. Problem is, that the planning is too far behind and the car cannot execute it successfully.

Travelling 36 mph, it's obvious that the car has botched the intersection and a disengagement is the correct action. Even if the car planned to stop at the stop line, it's going too fast at this point to stop without maximum braking.

I've had my car do something similar on an older 11.x. In my case, I was on a 55 mph rural road and came up on a perfectly visible stop sign with no occlusions. Fortunately, there were no other cars. My car put up the stop sign in the viz and displayed the stopping message. But it did both way too late. Since there was nobody anywhere nearby, I let the car do what it could and it hit the brakes way too late and blew through the intersection. It was a total failure - no excuses. It's why there's a human behind the wheel.
 
We don't know whether FSD planned to stop or not in this situation. If the car is at some distance before the stop sign then FSD should display a stop sign on the screen (similar to display the speed limit) to alert the driver. Then, if FSD intends to stop then it should display another bigger and glowing stop sign on the screen to tell its state of mind. If the driver does not see the bigger stop sign then he/she should take over. This will help the driver feel more comfortable.

We don't need the message "Stopping for stop sign". It's not easy to see. An icon is worth a thousand words.
 
In a quick drive I couldn't find any blind stop signs to match this scenario, but when rolling the speed down as it approaches a stop sign I did get it to slam on the brakes at ~5 MPH right before the stop line once. I tried to repeat the exact same scenario and couldn't get it to do it again, all of the other stops were gradual and smooth.

In 22 miles of FSDb driving it only triggered the "hard braking" in my insurance app, that one single time. (Which is a large improvement from older versions.)
 
I think this was an embarrassing look for Ross and for FSDb. In real time that looked super scary and was definitely worthy of a sharp disengagement with full brakes. Automatic Emergency Braking should maybe have activated even given how close a call it ended up being.

I think something that others haven't mentioned is that there is no Chevron sign indicating car slowing down and there is no line indicator of where the car is going to stop. This is all around a clear example to me of the car relying on vision only and incomplete map data. It needs to see the stop sign to plan the stop. That intersection is super tough - completely blind from the right side so the only way you see the white car is after it's basically 75% through the turn. But the car should have been easing to a stop way way before that to make for a comfortable stop at a stop sign.

This was truly a failure of FSDb during an important "test" of a hyper critical individual with an axe to grind with Tesla. It will definitely make it to the mainstream media and paint all the good that FSDb has done in a really bad light for a little while.
 
Interesting video. It does seem he cut off FSD before it had the chance to stop, but it was going too fast in my opinion to not take over.
Even if it was a flawless drive, FSD isn't worth the money. Update after update my car still pulls blindly into traffic to turn at an intersection (because it can't see jack with the side cameras so far back). I am hopeful, but it doesn't seem to be getting better as fast as promised, and I doubt Tesla will fix what seems to be a gaping hardware issue with the cameras for all FSD subscribed cars.
 
I think something that others haven't mentioned is that there is no Chevron sign indicating car slowing down and there is no line indicator of where the car is going to stop.
FSDb had just put up the "Stopping for stop sign" message when Ross disengaged. The chevrons and stop line indicator probably would have been displayed moments after that had he not intervened. (It might have even slammed the brakes on its own about the same time.) It certainly should have been slowing earlier, and it was good that Ross did intervene.

It would be nice to get a bunch of people together during a slow time of day to test some scenarios at the intersection. Other people could stop and wait at the other stop signs as FSDb approached to make testing what FSDb would do safer.
 
The chime clearly tells us when FSD beta is disengaged. It is at time code 0:37 after the car has passed the stop sign and is at the stop line. There is no doubt. To claim FSD beta is disengaged before that is disingenuous.
if the situation was due to fsdb malfunction or if it was mode confusion on Ross’s part doesn’t really matter from a road safety perspective even though I strongly believe it was the former. I believe the phrase “Gaslighting“ would be appropriate here but…

Instead, let’s talk about the sign that warned about the upcoming stop sign at 0:26 which is 10s earlier, and the FSDb didn’t slow down at all...
 
Last edited:
  • Like
Reactions: QUBO and diplomat33
It is very evident that the stop sign is way behind the actual intersection. If you stop at the stop sign there is no visibility of the traffic, especially from the right side from where the white car came. Evidently FSDb planned to stop beyond the stop sign such that it maintains visibility of the traffic
 
  • Like
Reactions: MP3Mike
I think it is still too reliant on map data. I have the opposite problem: stopping at intersections that used to have stop signs…clearly they’re hesitant to fully trust their vision stack.
Another way to test this theory, we use Google Maps history feature to see if that stop was recently installed.

They put up a stop in my neighborhood recently and FSDb will stop for it but in previous versions it did not. In either case it still behaves differently around this stop than others. It’s like for most stops it knows they exist and so it begins braking at the right time. For this new one, it brakes later (and in previous versions not at all). I reported that error many times and eventually it was fixed.
 
WOW. FSD Beta not just flew past the stop sign and almost crashed head on into a car that was making a turn in front of the Tesla. It is only Ross' quick intervention that prevented a major crash. It is just one example of how far behind Tesla is on FSD. it might "work" everywhere but it is nowhere near safe or reliable enough to remove driver supervision. Tesla FSD beta is far behind Waymo and Cruise.
And if the theory of a new a stop sign is right, then ask yourself: if Waymo were allowed to drive this fast on a premapped road but they had premapped prior to the stop install, would it do the exact same thing?

This issue to me highlights why Tesla’s tech is ultimately the more robust tech, since they will eventually be reliant on vision to react like humans to signage. (But vision is hard, though perhaps getting easier with recent breakthroughs from places like OpenAI.)
 
The chime clearly tells us when FSD beta is disengaged. It is at time code 0:37 after the car has passed the stop sign and is at the stop line. There is no doubt. To claim FSD beta is disengaged before that is disingenuous.
To be fair Dip, that chime sometimes plays a second or two after I disengage. The sound is a lagging indicator. (at least in my
mind it lags… we need to test to see by how much and with what variance)
 
  • Like
Reactions: MP3Mike