Another tragic fatality with a semi in Florida. This time a Model 3

[verygreen]

The problem with conventional radar is that it isn't accurate enough to differentiate obstacles from background objects. It just sees a large "blob" that could be a semi trailer or just an overpass

This is incorrect. Radar works wonders in those scenarios outlined. It also has a lot of smarts (many of which Tesla now disables because they think they can do it better).

It most certainly does not see "blobs" either. and it can tell you speed and the remaining smarts still tell you if the thing is moving and in which direction or if it was ever observed as moving while radar was observing it.

 

[seattlite2004]

Graphic of the vectors:

 

[chronopc]

I agree, but Tesla has been misleading people for 3 years on this, and a lot of folks are all too willing to drink it up. And while some people have been pointing this out for a while, they get voted down into oblivion or insulted.

This often comes up when referring to FSD allowing people to lease out their cars and make money from it being a robo-taxi. Some people thought it was a 2-year time-frame, I know I said more like 10 - 15 years and got laughed at - but a robo taxi needs full autonomy in the vast majority of situations, with no intervention, whereas NoA and any non-highway equivalent, clearly relies on 100% attention from a driver.

What's frustrating (as far as any web-forum can be truly frustrating) is that in being skeptical about FSD's capability in the near term, doesn't diminish Tesla's advantage in the market place, which is very real. No one else is any closer than they are.

In terms of robo taxi, Waymo (Google) is closer. They've already released it in Arizona.

We spoke to a Waymo One customer about how robot taxis get confused by rainstorms

 

[diplomat33]

Graphic of the vectors:

View attachment 408543

Thanks. I think FSD would be able to handle this scenario because the front side cameras would definitely have seen the truck and measured that it was on a collision course and taken action before a collision occurred.

 

[Daniel in SD]

This is incorrect. Radar works wonders in those scenarios outlined. It also has a lot of smarts (many of which Tesla now disables because they think they can do it better).

It most certainly does not see "blobs" either. and it can tell you speed and the remaining smarts still tell you if the thing is moving and in which direction or if it was ever observed as moving while radar was observing it.

Then why does every radar cruise control system documentation have warnings that the system can't see stopped vehicles?

 

[Daniel in SD]

In terms of robo taxi, Waymo (Google) is closer. They've already released it in Arizona.

We spoke to a Waymo One customer about how robot taxis get confused by rainstorms

* with a safety driver. It's still functionally a level 2 system. I do agree that they're way ahead though.

 

[diplomat33]

In terms of robo taxi, Waymo (Google) is closer. They've already released it in Arizona.

We spoke to a Waymo One customer about how robot taxis get confused by rainstorms

I have to chuckle a bit at the title, "gets confused by rainstorms". Interesting how folks complain that Tesla can't do FSD in the rain, yet Waymo with their sophisticated sensors, including LIDAR, also don't doesn't do well enough in strong rainstorms either.

 

[chronopc]

* with a safety driver. It's still functionally a level 2 system. I do agree that they're way ahead though.

Waymo uses a mapping approach so in certain situations they are way ahead. But to be fully autonomous in all situations, I would say Tesla's approach is a better one in the long term. Not sure how long it would take for any system to not need a "safety driver" though. Two years? Maybe. Longer? Definitely.

 

[chronopc]

I have to chuckle a bit at the title, "gets confused by rainstorms". Interesting how folks complain that Tesla can't do FSD in the rain, yet Waymo with their sophisticated sensors, including LIDAR, also don't doesn't do well enough in strong rainstorms either.

You can't map rain.

 

[verygreen]

Then why does every radar cruise control system documentation have warnings that the system can't see stopped vehicles?

There are multiple possible reasons. The car vendors might have decided to disregard some of the stationary radar returns to cut down on number of phantom brakes as one of a more technical examples. Less technical examples could include a simple cya so in case something goes wrong you can always blame the user.

 

[Eno Deb]

This is incorrect. Radar works wonders in those scenarios outlined. It also has a lot of smarts (many of which Tesla now disables because they think they can do it better).

It most certainly does not see "blobs" either. and it can tell you speed and the remaining smarts still tell you if the thing is moving and in which direction or if it was ever observed as moving while radar was observing it.

Did you read the rest of the posting? Of course it can detect relative speed, which is precisely what it uses to filter out most stationary objects to avoid false positives. But the original point stands. Traditional automotive radar (typically at 24GHz) has a rough idea of the size and position of an object (and a much more accurate idea of the distance and relative speed), but does not have the resolution to differentiate stationary objects reliably, such as telling the side of a semi trailer from an overpass.

 

[Brian's ReX]

You can't map rain.

Have you never seen a map of SEATTLE???

 

[electracity]

LIDAR would have less ability to spot this than a camera, unless the LIDAR was mounted on the roof, because what allows the camera to "see" this is the relative motion between the car in front and the object in front of that, via the wide-angle lens on the top of the windshield.
.

LIDAR would see a sold object in the path of the vehicle, regardless of the mounting height. It simply measure distance.

Not that I want to get into a LIDAR debate. I have no doubt that eventually just vision will be good enough. But I suspect the first autonomous vehicles will have robust multiple sensor types. Tesla has one robust sensor type.

 

[chronopc]

Have you never seen a map of SEATTLE???

We get more rain now than Seattle.

San Francisco has seen nearly twice as much precipitation as Seattle this year

 

[verygreen]

but does not have the resolution to differentiate stationary objects reliably, such as telling the side of a semi trailer from an overpass.

it's not the work of radar to tell the two apart (but trailer has an important property - it tends to be moving sideways). Radar does give you a heading so by combining that with distance you have a good idea how high this something is.

Based on my experiments radar does a good job showing stationary objects, both the background ones and obstacles. The problem is how do you determine which ones ar eobstacles and which ones are not. This is not radar task though. Lazy people tend to filter out all stationary objects. Smart people come up eith smart solutions.

Tesla has semi-dumb solution of "if it's stationary but it's on our map of stationary objects, we'd ignore it otherwise we'd brake for it" and "if it's moving sideways we'd brake for it even if it's just a freight train moving on a railway bridge that IS marked as a stationary object on our map"

 

[Brian's ReX]

We get more rain now than Seattle.

My apologies!

 

[ranss12]

Seems to me that the truck driver is at fault here - failure to yield right of way. Why are there "no charges pending"?

 

[Eno Deb]

it's not the work of radar to tell the two apart

But that was the original question I replied to. If you fuse it with input from other sensors or map data that improves things of course.

Tesla has semi-dumb solution of "if it's stationary but it's on our map of stationary objects

What map data is that and how is it obtained?

and "if it's moving sideways we'd brake for it even if it's just a freight train moving on a railway bridge that IS marked as a stationary object on our map"

Well, in this case it apparently failed (perhaps because the lateral speed was too low to pick up the movement at radar resolution?).

 

[VT_EE]

The problem with conventional radar is that it isn't accurate enough to differentiate obstacles from background objects. It just sees a large "blob" that could be a semi trailer or just an overpass. Automotive radar applications avoid this problem by ignoring anything that doesn't seem to move along with the car (the relative speed can be detected by Doppler shifts of the reflected radar signal). So essentially radar is good at detecting other cars, but not obstacles that are stationary or moving perpendicularly in front of the car.

Exactly! That truck looks like a bridge to the radar. I don’t think Tesla implemented the ability to detect a stopped vehicle using the cameras until VERY recently. Remember the case where you approach a stop light with a vehicle already stopped at the line. The radar never sees it moving and the Tesla would just plow into it. You need the camera as the primary sensor in the tractor trailer crossing the road scenario. I doubt this poor person’s car had it.

 

[Daniel in SD]

Seems to me that the truck driver is at fault here - failure to yield right of way. Why are there "no charges pending"?

Because if we filed charges every time someone was forced to cede their right of way 99% of drivers would be convicts.
How many drivers can honestly say that they have never taken someone else’s right of way?
Also, legally you’re required to avoid accidents even if you have the right of way.