Phantom braking will get a lot worse before it gets better

[mswlogo]

@verygreen can perhaps show a live feed video when it's phantom braking and what it sees.

Phantom braking the later versions has been horrible!!! What exactly is going on? What does it see ?

Longer days, more shadows.
Summer Solstice is just a few weeks away

I don’t think Phantom braking has changed since it first surfaced. It’s so environment sensitive.

 

[EVNow]

It clearly was spooked by oncoming traffic in my case, but those cars are not marked as obstacles or anything like that.

That’s my experience as well. Oncoming traffic spooks the car sometimes.

I wonder whether they made detection of collision more sensitive (i.e. reduced threshold probability values) because of some well publicized accidents.

 

[conman]

I wonder whether they made detection of collision more sensitive (i.e. reduced threshold probability values) because of some well publicized accidents.

Both the phantom braking, and the more frequent nags did seem to correlate chronologically with the backlash from those accidents.

 

[emmz0r]

Longer days, more shadows.
Summer Solstice is just a few weeks away

I don’t think Phantom braking has changed since it first surfaced. It’s so environment sensitive.

How can the radar pick up shadows? @verygreen said it didn't spot anything visual or tag anything

 

[mswlogo]

How can the radar pick up shadows? @verygreen said it didn't spot anything visual or tag anything

It’s not the radar triggering phantom braking. It’s the “traffic aware” cameras.

 

[rnortman]

The rest is history. Its one of those things where necessity has become a virtue.

Other than saving cost, what is the virtue?

 

[emmz0r]

It’s not the radar triggering phantom braking. It’s the “traffic aware” cameras.

But @verygreen said that they didn't detect anything.

 

[mswlogo]

But @verygreen said that they didn't detect anything.

I have not kept up with that conversation. I doubt it looks that “wide” and far to account for some of its wonky behavior if it was all radar. It’s hard to know for sure if it’s shadows or the objects making the shadows that confuses it.

I personally find it phantom brakes less at night. But it has happened at night too. When that has happened I think it’s sometimes headlights from other cars coming up along the side of you. I find it nearly impossible to predict when it will phantom brake.

 

[EVNow]

Other than saving cost, what is the virtue?

You actually get to put your hardware in 400k cars and collect training data. Can you imagine putting that ugly Lidar on top of S or 3 ?

Cruise proudly says they do 1,400 unprotected left turns a day. Together Teslas are probably doing 400,000 unprotected left turns a day - if each car does even one per day. NN needs lots of data.

 

[Eno Deb]

You actually get to put your hardware in 400k cars and collect training data. Can you imagine putting that ugly Lidar on top of S or 3 ?

Cruise proudly says they do 1,400 unprotected left turns a day. Together Teslas are probably doing 400,000 unprotected left turns a day - if each car does even one per day.

For how many of those turns does Tesla receive detailed data, including highly accurate location information (to determine the precise trajectories and lanes the drivers chose), oncoming traffic and pedestrians that prevented them from making the turn etc.?

NN needs lots of data.

And yet, Cruise has demonstrated that they can reliably do it autonomously and Tesla has not.

 

[EVNow]

For how many of those turns does Tesla receive detailed data, including highly accurate location information (to determine the precise trajectories and lanes the drivers chose), oncoming traffic and pedestrians that prevented them from making the turn etc.?

We don't know. We just know they can get as much information as they want.

And yet, Cruise has demonstrated that they can reliably do it autonomously and Tesla has not.

Yet, Cruise has not done one mile of autonomous driving in a commercially available car !

I'd say neither have demonstrated anything until we can actually experience it. Afterall, Musk says unprotected left turns on his car are getting much better.

BTW, are you claiming all the data Tesla can collect is not of any use ? i.e. Tesla's data advantage is useless ?

 

[Eno Deb]

We don't know. We just know they can get as much information as they want.


Yet, Cruise has not done one mile of autonomous driving in a commercially available car !

I'd say neither have demonstrated anything until we can actually experience it.

I have experienced it, and so have many others.

BTW, are you claiming all the data Tesla can collect is not of any use ? i.e. Tesla's data advantage is useless ?

I believe the much-vaunted data collection is far less extensive than many believe. There are many factors other than the number of vehicles that limit the amount of telemetry that can actually be collected and processed to a point where it becomes useful training data.

 

[EVNow]

I have experienced it, and so have many others.

​

When you say "experienced", were you able to drive in a road of your choice without a Cruise guy in the car ? Comparable to the topic of this thread - phantom breaking experienced by many of us when driving using AP, all over the world ?

I believe the much-vaunted data collection is far less extensive than many believe. There are many factors other than the number of vehicles that limit the amount of telemetry that can actually be collected and processed to a point where it becomes useful training data.

​

Ofcourse the second part is true - but the first part is difficult to pin point. It is not clear how to quantify what others believe and what you believe to be the truth.

Apparently all these companies have 1000+ people in India doing the labeling. So, millions of labeled images per month. The question is how varied is the data ... we'll know in a year if Tesla is better there.

 

[Eno Deb]

When you say "experienced", were you able to drive in a road of your choice without a Cruise guy in the car ? ​

Yes. Starting in late 2017, they have occasionally sent out invitations for public tests in SF to the media and certain . We entered an address of our choice and the car took us there (took about 20-30 minutes). It wasn't perfect (the car was often overly cautious), but it handled things like unprotected left turns, jaywalking pedestrians and the usual inner-city chaos without a single intervention by the safety driver. Tesla is nowhere near from what they have publicly shown so far.​

 

[EVNow]

Yes. Starting in late 2017, they have occasionally sent out invitations for public tests in SF to the media and certain . We entered an address of our choice and the car took us there (took about 20-30 minutes). It wasn't perfect (the car was often overly cautious), but it handled things like unprotected left turns, jaywalking pedestrians and the usual inner-city chaos without a single intervention by the safety driver. Tesla is nowhere near from what they have publicly shown so far.​

Got it - so a demo ? You didn't actually drive it ...

Yes, definitely Cruise and Waymo are ahead in demos. Once Tesla released NOA on city with HW3 it would be a good comparison point.

 

[Eno Deb]

Got it - so a demo ? You didn't actually drive it ...

WTH are you even talking about? You don't drive autonomous vehicles, you ride them. And that's what we did. There was a safety driver, but he didn't have to intervene even once. And it was not a pre-planned route or anything like that. What more do you want?

Once Tesla released NOA on city with HW3 it would be a good comparison point.

Exactly. But I suspect they won't come out of that comparison looking very good.

 

[EVNow]

What more do you want?

A commercial car that does this in my city ?

Exactly. But I suspect they won't come out of that comparison looking very good.

Except for the above part ?

Anyway, I think we are OT now.

 

[Eno Deb]

A commercial car that does this in my city ?

I'm willing to bet that we'll see commercial Robotaxis powered by Cruise long before we see Tesla Robotaxis. Maybe if I see NoA managing to navigate busy traffic in downtown SF for half an hour without interventions that would change my opinion.

 

[electronblue]

Yes, definitely Cruise and Waymo are ahead in demos. Once Tesla released NOA on city with HW3 it would be a good comparison point.

I guess the bigger point here is: There seems to be legitimate reason to believe Cruise and Waymo are ahead. Period. In their case we have actual demos and even some riders and rides that we could probably chalk closer to actual use than only demos.

With Tesla we have only Elon’s word on their city driving. As far as I know the Autonomy Investor Day rides didn’t alllow free addresses and travelled in easier suburbia or countryside.

So reasonably it looks like Waymo and Cruise are ahead. Tesla has a story on how their trajectory will surpass others faster in general autonomous driving but so far that is a story not much more.

The latest 16.x updates don’t actually instill much confidence in me seeing how the adjacent cars are making all sorts of priouettes and turning sideways on the IC now. Reliably assessing objects in traffic is paramount to reliable city driving. Again the story that HW3 will make all this better is so far only a story.

 

[EVNow]

So reasonably it looks like Waymo and Cruise are ahead. Tesla has a story on how their trajectory will surpass others faster in general autonomous driving but so far that is a story not much more.

But Cruise and Waymo are geofenced and seem to be stuck in demo mode for too long.

Anyway, the diametrically opposite paths they have chosen makes it all the more interesting.

I’m fairly confident I will be driving with NOA on city roads in my 3 earlier than riding in Cruise/Waymo robotaxis I in my city.