Tesla.com - "Transitioning to Tesla Vision"

[Knightshade]

Anecdotal isn't super useful.

And likely there'll be SOME things the new system already does better- and other things worse.


Know how we'll have some legitimate cause to think TV is as good or better than cam+radar overall?

When they remove the additional restrictions only TV has right now.

(and the fact they're making the restrictions marginally less restrictive, 80 vs 75, doesn't change the fact it's still restricted compared to the +radar system)

 

[Daniel in SD]

Know how we'll have some legitimate cause to think TV is as good or better than cam+radar overall?

When they stop using radar on all cars. This seems inevitable, they’re not going to want to maintain two versions. Even that won’t prove it and there will be people who swear that it’s worse and people who swear that it’s better - just like every software update.

 

[mrkisskiss]

I honestly don't understand the decision to remove sensors. But, a few years ago I realised that unfortunately, Musk has done exactly what he was always advising against:

"Don't let perfect get in the way of better".

It's a shame they didn't just look at improving the general usefulness of Autopilot. E.g, adjusting for traffic speed signs in the UK. AP1 did this just fine in 2014. My Model 3 still doesn't do this, and we're half way through 2021.

To be honest, I felt they basically lost their way when they announced they weren't going to use maps as part of their driving tech stack. This idea that "humans drive with two eyes and no maps so it must be possible for an AI to do it" is such a flawed hypothesis. Humans drive extremely badly on complicated (i.e non-US), unfamiliar roads. Again, it's based from this idea that all driving is LA freeway driving - straight forward, stopped at lights, or turning right at an intersection.

With a map, maybe they could have done it... I defy any good driver to confidently nail Swindon's magic roundabout without any kind of prior knowledge -ha!

 

[Knightshade]

I honestly don't understand the decision to remove sensors.

Watch the recent Karpathy video- he not only explains it, he cites 3 specific examples from real-world vehicle data.

Radar was making performance worse in many situations.

It's a shame they didn't just look at improving the general usefulness of Autopilot. E.g, adjusting for traffic speed signs in the UK. AP1 did this just fine in 2014. My Model 3 still doesn't do this, and we're half way through 2021.

AP2+ has done this (in the US anyway, I know EU regulations has nerfed many features) for a good while now too.

Mobileeye had a patent on the feature so it did take a few years for Tesla to develop around it FYI.

To be honest, I felt they basically lost their way when they announced they weren't going to use maps as part of their driving tech stack.

They have never announced that.

They explicitly use maps- just not millimeter-accurate HD ones, because it's impossible to scale those accurately.... (hence why everyone using them is heavily geo-restricted in where they work at all)

If they didn't use maps NOA wouldn't ever know what lane to get into to exit the highway. Nor would the car be able to start slowing for lights or stop signs not currently visible. Both of which the system does among myriad other examples.

 

[powertoold]

I had high expectations for this vision-only transition, but Tesla has exceeded my expectations, considering the difficulty of what they've accomplished. Again, *no one* has achieved a vision-based feature that is remotely close (in terms of accuracy, precision, and recall) to Tesla Vision. Hell, I'd say that there is no vision-based app or software (in any industry) that is as good as Tesla's in predicting real-world features. Tesla Vision is probably approaching the 99.9%+ in accurately predicting real-world features at ~100-150m.

 

[Daniel in SD]

Hell, I'd say that there is no vision-based app or software (in any industry) that is as good as Tesla's in predicting real-world features.

 

[WhiteWi]

Has radar not vision only

 

[powertoold]

You should have posted this video. I had a good laugh at it:

 

[Daniel in SD]

Has radar not vision only

 

[DanCar]

Car path y discusses low visibility situations at 14:50 mark. A dust cloud and what he claims is heavy snow. I'm thinking heavy fog and heavy rain would be much worse than that dust cloud issue. That video of heavy snow doesn't seem like it would meet my definition of heavy snow, but sometimes difficult to judge weather from cameras.

 

[powertoold]

Car path y discusses low visibility situations: a dust cloud and what he claims is heavy snow. I'm thinking heavy fog and heavy rain would be much worse than that dust cloud issue. That video of heavy snow doesn't seem like it would meet my definition of heavy snow, but sometimes difficult to judge weather from cameras.

I was wondering about that as well. We'd have to wait for snow / rain season to see. There's decent performance in user videos where there's moderate rain mist behind lead cars. But we don't have many examples of heavy rain or heavy snow.

 

[Itsame]

I think it's obvious it was the chip shortage and cost savings to remove radar. He's trying to spin it as a positive. It's not.

 

[DanCar]

I think it's obvious it was the chip shortage and cost savings to remove radar. He's trying to spin it as a positive. It's not.

I'm on the fence. Agree chip shortage was the reason, but hopeful they can make it work. Advantage is lower cost of car and lower cost to repair.

 

[powertoold]

I guess this means all FSD-equipped cars are going vision-only in a month or 3 maybe definitely?

https://twitter.com/x/status/1409934915408478213

 

[Knightshade]

I guess this means all FSD-equipped cars are going vision-only in a month or 3 maybe definitely?

https://twitter.com/x/status/1409934915408478213

The "top rating for FSD computer cars" thing is interesting- I wonder how the folks who always insisted Tesla would never charge for better safety would respond to that.

 

[Daniel in SD]

The "top rating for FSD computer cars" thing is interesting- I wonder how the folks who always insisted Tesla would never charge for better safety would respond to that.

I interpret it to mean all HW3 cars. I never interpreted "never charge for safety" to mean free hardware upgrades.
If HW3 ends up providing better EAP performance and active safety feautures I wonder if Tesla will offer HW3 upgrades at a lower cost than FSD?

 

[powertoold]

If HW3 ends up providing better EAP performance and active safety feautures I wonder if Tesla will offer HW3 upgrades at a lower cost than FSD?

I vague remember Elon saying that Tesla would upgrade your car to HW3 if you subscribe to FSD, but then again, you're paying for safety.

I don't think Tesla themselves have claimed that they'll never charge for better safety

I'm pretty sure the transition of all FSD-equipped cars to vision-only is the precursor to The Button (tm).

 

[Knightshade]

I interpret it to mean all HW3 cars. I never interpreted "never charge for safety" to mean free hardware upgrades.

FWIW I agree- but when I raised the point a while back that eventually the AP/EAP codebase would stop improving for HW2.x cars while HW3 would continue to get better/safer- a bunch of folks insisted Tesla wouldn't ever put safety behind a paywall like that.

This seems to disprove their notion explicitly (though it'd been obvious for a while)

 

[DanCar]

https://twitter.com/x/status/1409879092531535882

Would be interesting if IIHS tested during rain, fog and snow conditions. In other words when visibility isn't great.

 

[mikes_fsd]

https://twitter.com/x/status/1409879092531535882

Would be interesting if IIHS tested during rain, fog and snow conditions. In other words when visibility isn't great.

Agreed, but only publish the results when they test -- say -- all 2021 models under the same conditions (so that there is baseline that is more accurate).