First Jerry rig everything, now mark rober 🧐

[Crissa]

The problem with the lidar/vision question isn't recognizing things : In almost all the cases of a serious problem it's not that the AI stack didn't see something, it's that it didn't react appropriately.

Cruze didn't get in trouble for not seeing the lady that was thrown under the car - it knew the lady was there, and made the wrong choice. (And the company failed to respond appropriately, either, humans lagging in their response to authorities).

The crash in Phoenix that killed a woman: The automatic braking system saw her, and so did the navigation system. The prior system had been disabled and the latter system decided the trash bags she was carrying weren't something to navigate around.

It's the decisions that's turned out to be difficult, not sensing shapes around the car.

Roads are often unclear on what's going on. Lines get painted wrong; signs are in the wrong place, corners are too tight or don't have enough vision for Humans and yet we let Humans drive on them!

Remember that barrier on 101 that a guy's Autopilot drove into almost a decade ago now? That rut had been there for years, I'd bounce off it in my car while commuting, too. And two (2!) Human drivers had slammed into it in the prior couple months, which is why the crash barrier wasn't there - along with the car behind the Model X which crushed him after it hit the wall.

It's the decision tree that's the hard part anymore, not the sensors.

-Crissa

Sponsored

 

[cybercricket]

Lidar is limited by glare the same way that vision is. It uses light waves, too.

The main difference is that cameras will typically be limited to capturing ambient light passively, limiting their ability to discriminate signal to noise. Reflections from the active light sources like lasers would only be limited by the safety limits, and would perform better even within those limits, especially since the detection at a fixed wavelength that isn't significantly impacted by the ambient light can be much more sensitive, further increasing the performance.

 

[Crissa]

The main difference is that cameras will typically be limited to capturing ambient light passively, limiting their ability to discriminate signal to noise. Reflections from the active light sources like lasers would only be limited by the safety limits, and would perform better even within those limits, especially since the detection at a fixed wavelength that isn't significantly impacted by the ambient light can be much more sensitive, further increasing the performance.

You do know that cars are required to have lights, right? And it's not illegal to put more out of visible band lights on a car? And cameras can take advantage of streetlights to see far in the distance?

-Crissa

 

[cybercricket]

You do know that cars are required to have lights, right? And it's not illegal to put more out of visible band lights on a car? And cameras can take advantage of streetlights to see far in the distance?

-Crissa

Let's be honest, sometimes you post dumb stuff. What has a better chance of reflecting from a solid in a fog and in the presence of ambient light and then registered by a sensor: broad (both shape and spectrum) beams from the headlights, or a narrow (again shape and spectrum) beam from a laser ? Also who is paying for installing streetlights in every shithole in America ?

 

[scottf200]

Many have and will pile on. It is human nature to join the crowd. Elon does have a pretty good track record of success. Often that success comes when the mindset of many are drifting from one idea to another. Elon is focused even when he appears not so.

Till then let people be the herd.
Buffalo Jump

Elon has drifted a LOT on "FSD". I recall presentations and him convincing "us" that Hi-Def maps were the way to go. I've been following Tesla AP/FSD closely from the beginning.

I think Tesla FSD will plateau and have various limitations. It is impressive in it's current state relative to the past and some other USA options from OEMs for personal vehicles but not in China and other places where Tesla FSD is run-of-the-mill. Tesla's displays use to be impressive but now they are ho-hum and China cars and others are better in accuracy and capability.

ASIDE: The current version of FSD has ridiculous speed controls. They appear to be black box controlled and are very annoying in my experience and many voices online. When you have to use a hack of adding 40% to try to trick it into doing something reasonable you know they are wrestling with it. Finally in the latest version they let you put a top limit. It gets reset regularly tho.

 

[scottf200]

can simulate lidar with depth perception exceeding humans.

Exceeding humans does not say much at all for depth perception. We are decent relatively close. Far off with/without speed and darkness (etc) is not great.

 

[65SoYoLO]

I thought (assumed yeah I know) that is was on FSD! That's an absolute BS test

 

[BannedByTMC]

Phil DeFranco interviews Rober about the video:

 

[CyberGus]

Looking at the stock chart, it seems like Luminar stock is going to be delisted.

This viral video might just be their last desperate attempt, don't you think?

A fall from $600 all time high to $5 per share can cause derangement. That might be the reason they are funding this disinformation.

Luminar seems to be doing well since the video was released...

:

 

[btcrealm]

Phil DeFranco interviews Rober about the video:

Thanks for posting that. I hadn't seen it.

I find it hard to believe he doesn't know that FSD just goes when you hit the button whether there is a destination or not. I do this all the time for driving down the highway.
I can also say that FSD and Autopilot do not react the same in similar situations. FSD is also a lot more smooth than Autopilot, especially at lights.
Maybe the cars react similarly in both modes but the cybertruck does not!

 

[BannedByTMC]

Thanks for posting that. I hadn't seen it.

I find it hard to believe he doesn't know that FSD just goes when you hit the button whether there is a destination or not. I do this all the time for driving down the highway.
I can also say that FSD and Autopilot do not react the same in similar situations. FSD is also a lot more smooth than Autopilot, especially at lights.
Maybe the cars react similarly in both modes but the cybertruck does not!

They act differently on my Model 3 as well.

 

[freyguy]

I am not sure if the vehicle needs LIDAR, but every camera needs to be self cleaning. They are doing the cybercab testing in Austin for a reason, the weather allows it. In places it snows regularly, you would have to shut down the fleet for 4+ months a year.

I personally use FSD about 70% of the time in the summer and 10% of the time in the winter. Cameras are occluded, speed is limited. You genuinely would need to pull over every 30 seconds in the winter to keep the cameras clean, FSD is unusable in bad weather conditions.

I agree with the self cleaning, but for now we now get a camera visibility section in the service menu with the latest update. To at least inform us there is an issue. I plan to intentionally cover 1 and see how quickly I get a notification.

 

[65SoYoLO]

Thanks for posting that. I hadn't seen it.

I find it hard to believe he doesn't know that FSD just goes when you hit the button whether there is a destination or not. I do this all the time for driving down the highway.
I can also say that FSD and Autopilot do not react the same in similar situations. FSD is also a lot more smooth than Autopilot, especially at lights.
Maybe the cars react similarly in both modes but the cybertruck does not!

and you can see he is turning the wheel slightly which will cut off FSD.

 

[REM]

Phil DeFranco interviews Rober about the video:

One thing I would like everyone to take notice about the hit-piece video, is that Mark attempted to engage auto-pilot 2 times while quickly approaching the "wall", and twice in a row it denied him. It probably did so because he was constantly accelerating manually with his foot.

It was only on the 3rd attempt that it accepted control over to the drive computer, and by that time I estimate he was less than 100ft away from the prop-wall. Around the 50 ft mark, it give control back over and starts to warn him. In other words, it's very likely that he found a niche edge case where he could seemingly get the AP to engage briefly before smashing into the wall.

Why didn't Mark engage AP half a mile away? Is it because he tried that, and AP stopped for the obstacle? It's plainly obvious that he purposefully found this edge case and ran with it.

Tesla has all this telemetry on hand. And I hope they publish it for everyone to see.

He also mixed different video takes in the final cut (some done previously on a different day, apparently). In this interview he's pretending to be daft about the difference between AP and FSD. Intellectual fraud at best; complete fraud FUDster at worst. I sincerely hope Tesla perma-bans his access to any amount of automated driving.

 

[Sjohnson20]

Somebody make this wall and try it with FSD on the Cybertruck!

Sponsored