Autonomous Car Progress

[Doggydogworld]

What's there to "believe"? It showed clearly in the video what happened.

And the NBC Dateline 1992 story on clearly showed a GM pickup truck exploding when struck in the side by a car. And Top Gear viewers clearly saw 4 guys pushing a Tesla Roadster to the garage. Video doesn't lie, but liars sometimes shoot video.

Are you suggesting the reporter faked it?

At first I would have said no, but now I'm convinced of it. It's trivial to do:
1. Hail the car, walk a few hundred feet and start the cameras. "Why is it way over there?"
2. Put something in front of car at a red light, roll cameras after the light turns green
3. Move the drop-off pin somewhere in Corona Heights Park that's closer to State Street than Museum Way
4. Act exasperated and hang up on the patient customer support lady who tries to help the "confused customer" move the pin
5. Say "I'll just hit confirm and see what happens", knowing exactly what will happen

None of this proves they rigged it. But there are hundreds of Waymo ride videos out there. Once every 50 or so one has a minor problem. And this lady gets four major problems on her very first trip? That doesn't strike you as a little bit suspicious?

Why does Waymo take Maya straight to the Randall Museum, but not exasperated journalist lady?

 

[Bladerskb]

What are you talking about? FSD Beta is more than capable of reading a closed road sign and rerouting. Here's a recent example at 9:15, and it does it way faster than Waymo:

Lol even dirtyTesla knew it was a fluke (and more to do with the angle and direction the car was coming from and not that it read that the road was closed) as FSD Beta has consistently failed at road closures let alone partial road closures.

you can also see it CLEARLY with this video (at 6:45) from the latest build.
Even the superfan said "I wouldn't necessarily call that a regression because the Beta has never been good at reading signs"

If you're referencing reading hand-held stop signs, there's nothing inherently wrong with Tesla's stack that would prevent reading and obeying them, so it must be a choice Tesla has made. If they can identify roadside stop signs, they can also identify hand-held ones. Maybe they didn't want anyone with a printer and some red ink to be able to permanently stop any car with FSD Beta.

Oh look the patented predictable Tesla fan logic shows up.

If Tesla doesn't have it then its useless. If they do then it makes them 10 years ahead.
FSD beta doesn't support highways? Oh that's cause highway are pointless. FSD Beta finally supports highway? Waymo doesn't even work on highways that shows you how Tesla is 10 years ahead.

Tesla has forward radar which is all you need. Tesla removes forward radar? That's because radar sucks and vision is all you need. Tesla adds back radar? Tesla leapfroged everyone to make the first high res radar making them 10 years ahead.

I could go on and on and on.
The Tesla logic never fails to amuse me.

 

[willow_hiller]

Oh look the patented predictable Tesla fan logic shows up.

Explain exactly what it is about Tesla's FSD approach that precludes them from reading and reacting to hand-held stop signs.

Would more sensors help? Predictable Waymo employee logic.

 

[Bitdepth]

Explain exactly what it is about Tesla's FSD approach that precludes them from reading and reacting to hand-held stop signs.

That's not what @Bladerskb is referring to. They are talking about following traffic direction when someone is directing traffic using gestures. Based on all the presentations Tesla has given, they have not talked about the software being able to recognize gestures and reacting to it accordingly. Elon Musk has talked about it being "soon" able to recognize gestures back in 2021.

https://twitter.com/x/status/1414435086548733952

“It will soon capture turn signals, hazards, ambulance/police lights & even hand gestures.”

Does it currently have that capability, evidence strongly suggests it does not.

Would more sensors help? Predictable Waymo employee logic.

Yes, more sensors help, if you look at research papers, human pose estimation using multimodal sensors perform better than lidar or camera alone. It can be done with camera alone but lidar helps, think using just camera vs ToF type sensor for skeletal tracking. Camera alone has come a long way since way back when. It's pretty easy to tell FSD beta is not taking human poses into consideration when tracking VRU. I'll use the video above.

If we look at the video, you can see it is tracking the VRU well, their general direction and flow is tracked but their pose is wrong. The direction they are facing is not being "tracked", my guess would be it is based on the direction the whole body is moving and not the pose. You can see the pose when they stop moving changes to face the wrong direction and their body pose is not tracked. Even when they stand still, the visualization shows them animating.

If that's not convincing, then perhaps Elon Musk would.

We need to predict what pedestrians will do based on their behavior, including limb angle & direction of sight. FSD currently sees all pedestrians as cuboids, so is overly cautious.

https://twitter.com/x/status/1644942345383673856

Let's use Waymo for example because they are clearly at the forefront of ADS.

In 2019 they were demonstrating their software being capable of following traffic control police officer's gestures.

They have talked extensively about how they try to solve this issue using not only hardware capable of sensing nuance but state of the art machine learning as well.

Multi-modal 3D Human Pose Estimation with 2D Weak Supervision in Autonomous Driving

 

[Dan D.]

“It will soon capture turn signals, hazards, ambulance/police lights & even hand gestures.”

https://twitter.com/x/status/1414435086548733952

Does it currently have that capability, evidence strongly suggests it does not.

It is indeed easy, and optimistic for claims to be made that Tesla COULD do something, or might be able to do it very soon. But what matters is whether they actually are doing it. Reading and reacting to all the required signs, and correctly interpreting VRU behavior, and all the other extra things involved with driving that Tesla is not currently doing is so taxing that I'm not sure if they have the hardware required to handle it.

Also, I don't know why we always have to guess what Tesla is currently doing, more transparency in their program would be helpful. I suppose it's better from a PR sense not to advertise all the things it doesn't do, and just focus on the things it does do.


Also, +1 for capturing the Scoot Coupe rentals!

 

[willow_hiller]

That's not what @Bladerskb is referring to. They are talking about following traffic direction when someone is directing traffic using gestures.

That was an awful example of following gestures, if that's actually the original point and you're not just changing the topic. The Waymo driver stops dead for a full 30 seconds, and then takes the only available route forward after a guy points to the left a few times. Would it have driven over a curb if he had pointed to the right? Probably not; following hand gestures was secondary to taking the only available route around.

FSD Beta is pretty good at finding the intended route of traffic through cones and markers, and that's a much more useful ability than watching hand gestures. And it's capable of doing so without an internet connection. For all we know, the delay in Waymo taking action in that video was caused by it waiting for information from a remote operator.

 

[stopcrazypp]

And the NBC Dateline 1992 story on clearly showed a GM pickup truck exploding when struck in the side by a car. And Top Gear viewers clearly saw 4 guys pushing a Tesla Roadster to the garage. Video doesn't lie, but liars sometimes shoot video.

ABC7 isn't NBC Dateline. Top Gear is known for pulling stunts. Local stations don't have the resources to go to those extents. I have interacted with them before, they have about half a day to film a story and get it out.

At first I would have said no, but now I'm convinced of it. It's trivial to do:
1. Hail the car, walk a few hundred feet and start the cameras. "Why is it way over there?"
2. Put something in front of car at a red light, roll cameras after the light turns green
3. Move the drop-off pin somewhere in Corona Heights Park that's closer to State Street than Museum Way
4. Act exasperated and hang up on the patient customer support lady who tries to help the "confused customer" move the pin
5. Say "I'll just hit confirm and see what happens", knowing exactly what will happen

None of this proves they rigged it. But there are hundreds of Waymo ride videos out there. Once every 50 or so one has a minor problem. And this lady gets four major problems on her very first trip? That doesn't strike you as a little bit suspicious?

Why does Waymo take Maya straight to the Randall Museum, but not exasperated journalist lady?

As far as I can find, it's not possible to halt a Waymo and get it show that message simply by blocking its way.
Here's an example of a police blocking a Waymo and according to Waymo's statement the car was able to execute a U-turn within about 90 seconds. The video never shows the message that was shown in the reporter segment.

Also, as Waymo did in previous cases, if they faked it, Waymo can easily make a statement about what happened. As such I find it unlikely they faked it. It's way too easy for Waymo to disprove.

As for Randall Museum, did the person directly enter Randall Museum in the app? Were they going from the same direction? Did Waymo adjust the routing before or after report came out? Are there daytime vs nighttime restrictions on access to certain areas? The video only shows an end point but doesn't show the progress, so there is no way to know if the test was done under the same circumstances.

 

[Madmolecule]

I tried FSD yesterday. when it started raining I received a message about the FSD not working fully. Will the robot axis pull over during precipitation. Maybe that is why the auto-wipers don't work, as the car was not made for percipitation, or maybe I am in the fog.

Odd with "Vision Only" only one camera has a wiper? Still no display of camera quality or how dirty they are. I feel this would be important for a FSD vehicle.

 

[JB47394]

Will the robot axis pull over during precipitation.

We'll know that when somebody says that they have created a robotaxi and it drives through the precipitation that you have in mind.

Maybe that is why the auto-wipers don't work, as the car was not made for percipitation, or maybe I am in the fog.

In what way do the auto wipers not work for you?

Odd with "Vision Only" only one camera has a wiper?

There are three cameras in that housing, and they are the most critical. Please explain why you consider the current configuration as "odd".

Still no display of camera quality or how dirty they are. I feel this would be important for a FSD vehicle.

Why do you feel this would be important?

 

[Knightshade]

I tried FSD yesterday. when it started raining I received a message about the FSD not working fully. Will the robot axis pull over during precipitation.

Since RTs aren't actually a thing yet with this system one can only guess.

That said, FSDb has worked to a far greater degree in rain, even when getting the "degraded" message on screen, as compared to the older non-fsdb software, so it remains to be seen how it'll work out.

Odd with "Vision Only" only one camera has a wiper?

No, 3 of them do in HW2 and 3 cars. 2 of them in HW4 since they replaced the 3 behind-the-windshield cameras with 2 higher resolution ones.

Still no display of camera quality or how dirty they are. I feel this would be important for a FSD vehicle.

Why would displaying on screen camera quality info in a vehicle intended to not have or require a human present be important?

 

[diplomat33]

We'll know that when somebody says that they have created a robotaxi and it drives through the precipitation that you have in mind.

I tried FSD yesterday. when it started raining I received a message about the FSD not working fully. Will the robot axis pull over during precipitation. Maybe that is why the auto-wipers don't work, as the car was not made for percipitation, or maybe I am in the fog.

Tesla FSD is not good enough for autonomous driving in rain. That's why it displays the message that it is not working properly. A Tesla "robotaxi" would likely need to pull over. But Waymo has robotaxis that work autonomously in rain.

 

[Madmolecule]

We'll know that when somebody says that they have created a robotaxi and it drives through the precipitation that you have in mind.

I was told by Elon eons ago that I would be making $30K a year of my robotaxi that I purchased in 2018. just waiting on my software upgrade button. I am an early adopter so I will get it first, Elon said so

In what way do the auto wipers not work for you?

vision only wipers have never functioned as will as my 2012 BMW that has a sensor.

There are three cameras in that housing, and they are the most critical. Please explain why you consider the current configuration as "odd".
Vision only not good with dirty glasses. does not have other sensors like radar, lidar or ultrasonic to verify what its eyes are seeing. Since Tesla only uses data to tuned the model and does not use the data for prevision so it does not know what to expect and relies only on vision. sponsored by the tire replacement guys, pothole detection.

Why do you feel this would be important? either Tesla cannot compute the multiple sensors with current or future computers or they are to cheap. Vision Only uses only one sense or nearly senseless. check out this sub $100 presence sensor I am testing out my house, pretty amazing tech.

 

[Knightshade]

Tesla FSD is not good enough for autonomous driving in rain. That's why it displays the message that it is not working properly.

Except my FSDb car keeps driving in the rain- even with that message displayed- it does not ask me to take back over...even continuing to make turns at intersections and whatnot.... (this contrasts to old NoA code where it would drop back to basic AP in moderate rain and be unable to do anything other than single-lane/speed keep)

So I'm curious what special insight you have on the degree to which performance is limited by rain you can confidently say it's not good enough in rain and would need to pull over?

 

[diplomat33]

Except my FSDb car keeps driving in the rain- even with that message displayed- it does not ask me to take back over...even continuing to make turns at intersections and whatnot.... (this contrasts to old NoA code where it would drop back to basic AP in moderate rain and be unable to do anything other than single-lane/speed keep)

So I'm curious what special insight you have on the degree to which performance is limited by rain you can confidently say it's not good enough in rain and would need to pull over?

It's the difference between driving with supervision versus without supervision. The fact that FSD beta will display a message saying performance is degraded implies that it is not good enough to be a (driverless) robotaxi. Put another way, if FSD beta had the reliability needed to drive autonomously without supervision in the rain, it would not need to display a message that performance is degraded. FSD beta works now but that is because FSD beta requires a human driver as back-up, so the performance standard is lower. It can be degraded and still seemingly work just fine since the human is expected to take over if there is a problem. But in a (driverless) robotaxi, there is no human back-up, so if there is degradation in the performance, the robotaxi will likely need to pull over for safety reasons. Think of this way. Maybe in rain, FSD beta performance is degraded to say only 95%. That's good enough to drive around with driver supervision, it is not good enough to be a (driverless) robotaxi. Of course, the message does not say how much degradation. Maybe the degradation is only 2%. We would not notice any difference in driving, hence why it seems to drive just fine even with the degradation alert. But a 2% degradation would mean that it could not operate safely as a (driverless) robotaxi since to be a robotaxi, it needs to be 99.999%.

 

[JB47394]

I was told by Elon eons ago that I would be making $30K a year of my robotaxi that I purchased in 2018.

Stop listening to the promises of salesmen.

vision only wipers have never functioned as will as my 2012 BMW that has a sensor.

So they work, but you want them to work better.

Vision only not good with dirty glasses. does not have other sensors like radar, lidar or ultrasonic to verify what its eyes are seeing.

You should probably buy a Waymo or Cruise vehicle. I understand that they combine a number of different sensors.

Vision Only uses only one sense or nearly senseless. check out this sub $100 presence sensor I am testing out my house, pretty amazing tech.

Perhaps you should contact Tesla's FSD team and present your use and business case. I doubt they read these forums.

 

[Knightshade]

It's the difference between driving with supervision versus without supervision. The fact that FSD beta will display a message saying performance is degraded implies that it is not good enough to be a (driverless) robotaxi.

I don't think it does though, and just saying so doesn't make it so.

The only thing it implies is... performance is degraded.

That might be true in lots of ways that are NOT indicating what you suggest.

As a most obvious example maybe it means top speeds must be reduced as it can't see as far well enough- but below the lower speeds it's still safe.

Put another way, if FSD beta had the reliability needed to drive autonomously without supervision in the rain, it would not need to display a message that performance is degraded.

Except, again, it would.


A vehicle can have multiple ODD limitations even at L4.


"ODD L4 max 85 mph in clear visiblity"
"ODD L4 max 45 mph in heavy rain"

Both can be true and the vehicle still be L4....even though overall capability is "degraded" in the second case.


I certainly don't know the DEGREE to which rain degrades the systems performance.... but the point is- neither do you

. Think of this way. Maybe in rain, FSD beta performance is degraded to say only 95%. That's good enough to drive around with driver supervision, it is not good enough to be a (driverless) robotaxi.

Again this appears to be assuming facts not in evidence.

95% might very well be "good enough" depending what you lose in that 5%. If it's a lower top speed that's totally fine. And that's not the only possible example just the simplest.

.
But a 2% degradation would mean that it could not operate safely as a (driverless) robotaxi since to be a robotaxi, it needs to be 99.999%.

You are conflating "2% reduction in total capabilities" with "2% reduction in how rarely it needs a human"

Those are vastly, vastly, different things.

 

[diplomat33]

You are conflating "2% reduction in total capabilities" with "2% reduction in how rarely it needs a human"

Those are vastly, vastly, different things.

No, I am not conflating those two things. But perhaps, I should clarify that when I say degradation, I am referring to degradation in perception. Admittedly, that is an assumption but I think it is logical. We know rain makes it harder to see. The only sensors that Tesla uses are cameras. We know rain will cause cameras to lose some reliability. So the message "inclement weather detected - performance may be degraded" is likely a reference to the cameras not seeing as well due to the rain. And perception degradation would affect all driving tasks. And we know that even a small degradation in perception would mean the driving is potentially less safe and therefore the system might need to pull over. Can we say 100% that the vehicle would pull over? Of course not. But it is logical. And considering that Tesla FSD requires driver supervision even when the system is not degraded, it would surely need supervision when the system is degraded due to rain.

 

[pilotSteve]

So when do we expect Tesla to retrofit radar to all its FSD vehicles?

 

[Knightshade]

No, I am not conflating those two things. But perhaps, I should clarify that when I say degradation, I am referring to degradation in perception. Admittedly, that is an assumption but I think it is logical. We know rain makes it harder to see. The only sensors that Tesla uses are cameras. We know rain will cause cameras to lose some reliability. So the message "inclement weather detected - performance may be degraded" is likely a reference to the cameras not seeing as well due to the rain.

Ok, so far so good, but does not support your actual claims.

And perception degradation would affect all driving tasks. And we know that even a small degradation in perception would mean the driving is potentially less safe and therefore the system might need to pull over.

And here is where you jump to an unfounded conclusion.

I even gave examples of situations where part 1 would not lead to part 2... instead it might lead to a slightly narrower ODD in which it can continue to operate perfectly safely at L4.

 

[zoomer0056]

This is one area where a change in mindset will be required. We are so used to interacting with other human drivers where we can give a hand signal or head nod to communicate intent.

Mindset and hand gestures to an on coming computer driven car is not a problem. It's the computer that can't be trusted. Inside or outside the car. Just read all the comments from FSD Beta and other robotaxi stories.