Profound progress towards FSD

[bd7349]

So, limited public release in 6-10 weeks puts us between September 25th and October 23rd. That’s also right when Tesla usually releases their big yearly update (V9 and V10 were both out kn September) so it’s likely V11 will continue the trend.

interesting comment from Elon regarding how it’ll handle roundabouts:

Not perfectly at first, but yes. Will take maybe a year or so to get really good at roundabouts worldwide. The world has a zillion weird corner cases.

I for one am very excited to see how the new FSD system will learn and get better over time. With the Dojo training computer that they’re working on it should theoretically get better very quickly.

 

[diplomat33]

New tweet from Elon: https://twitter.com/elonmusk/status/1294374864657162240?s=20

Speed bumps and potholes too: https://twitter.com/elonmusk/status/1294376617960062976?s=20

I believe that it will be a quantum leap from what we have right now. Now, whether it will real reach full autonomy is another matter.

I am very happy to hear that Tesla is labeling potholes because I have a pothole on my daily route that forces me to disengage AP.

 

[ColonelRHS]

I would like to share something from BARRON's business news today. I will quote just some of the article whose lead headline was the Musk was becoming more like Henry Ford: " Two years ago after taking apart a Tesla Model 3, I couldn't believe how bad the body was put together says Sandy Munro, a former Ford Motor engineer whose consulting firm, Munro & Associates, specializes in reverse engineering and competitive analysis for the auto industry.
Everything else blew me away he says referring to Tesla's electric power train.
Now, after poring over every inch of a disassembled Model Y, its newest product, Munro says Tesla's improvement is remarkable. They're going to go from worst to first in a short time."....manufacturing expert Sandy Munro

 

[jeremymc7]

So, limited public release in 6-10 weeks puts us between September 25th and October 23rd. That’s also right when Tesla usually releases their big yearly update (V9 and V10 were both out kn September) so it’s likely V11 will continue the trend.

interesting comment from Elon regarding how it’ll handle roundabouts:



I for one am very excited to see how the new FSD system will learn and get better over time. With the Dojo training computer that they’re working on it should theoretically get better very quickly.

The one roundabout in my area was just redone (again) like it is every two years or so due to contact accidents.

There are now dedicated lanes with solid and dashed lines, turn arrows, and those white stick cone things.

So MANY this will be one of the easier roundabouts?

I think the city went overboard just so they can clearly say who was at fault depending on what lane they were in and when / how they moved their car. Now the city isn't (as much) at fault due to the poor design.

FYI the roundabout is a major roadway (Pacific Coast Highway) so it's a huge ton of traffic that goes through it.

 

[jeremymc7]

I believe that it will be a quantum leap from what we have right now. Now, whether it will real reach full autonomy is another matter.

I am very happy to hear that Tesla is labeling potholes because I have a pothole on my daily route that forces me to disengage AP.

Agreed but I wish it was the active system that lifts the wheel itself like some other manufacturers. If it can't be avoid you don't need to panic stop to slow down. Not to mention I've gotten SEVERAL blowouts at 5 mph so slowing down doesn't fix everything. Besides in heavy traffic it might cause more accidents.

 

[diplomat33]

So many companies have gotten way past the point of "almost zero interventions" for 10-20 miles years ago and they still don't have FSD. I guess the argument is that the methods that Tesla is using to get there is going to enable them to get to the ~100,000s of miles between failures required for FSD.

Yeah, the "Tesla argument" is that Tesla's large fleet collecting large and diverse data will allow Tesla to improve the disengagement rate much faster than other companies. I guess we shall see when the rewrite comes out. The exciting part is that we get the software in our cars so we can test it ourselves. So we can find out for ourselves how quick the progress is. I love numbers and stats. So I would love Tesla to release disengagement data once their "feature complete 4D rewrite" is released to the fleet. Then, we could plot the disengagement rate over time and quantify the progress. Elon claims FSD is improving at an exponential rate. Plotting the disengagement rate over time would be a nice way to visualize it and see if it is true or not. Maybe we could collect disengagement data from owners and do our own analysis?

But like you said, "almost zero interventions for 10-20 miles" is really bad. It sounds like the rewrite might give us self-driving that is starting pretty low in terms of reliability. We will then see how quickly it improves after that and whether the "fleet data" argument really holds water or not.

 

[jeremymc7]

Now, after poring over every inch of a disassembled Model Y, its newest product, Munro says Tesla's improvement is remarkable. They're going to go from worst to first in a short time."....manufacturing expert Sandy Munro

If they can improve their service and customer service areas, yes. Also consistency in build quality from unit to unit. I still think he needs to go back to doing everything you can with robots, but I understand why he back-paddled on this.

 

[jeremymc7]

Anyone want to throw the LiDAR monkey wrench into the "discussion" ?

 

[diplomat33]

Anyone want to throw the LiDAR monkey wrench into the "discussion" ?

It's still too early to rule lidar out. We know that lidar works really well for self-driving since companies with lidar have already achieved L4. But we have not seen what Tesla can do yet with this rewrite. Maybe the rewrite will be truly amazing self-driving or maybe it will still fall short of reliable self-driving. We don't know yet.

Mobileye did demo self-driving without lidar. So it is possible. But Mobileye still plans to add lidar for added reliability before they remove driver supervision. So it is possible that lidar is not strictly required to do self-driving but is required or desirable depending on what type of self-driving you want.

 

[diplomat33]

This is interesting.

 

[scottf200]

Via FB Tesla Tips & Tricks / Denis Y.

https://i.imgur.com/lXPPbAE.jpg

https://i.imgur.com/SK3DeMV.jpg

https://i.imgur.com/noUxuhy.jpg

 

[Bladerskb]

Thanks. In other words, Tesla is doing pretty standard machine learning but acting like it is some amazing radical breakthrough that will "solve FSD"?

Yes everything elon hypes up and calls a "quantum leap". Are all standard operating procedure. RNN & LSTM are all standard in autonomous driving software. Infact Tesla is the only one without a prediction network. The only thing they had was a CNN that did cut-off prediction. That was it. Unlike Self driving 101 where your car actually need to know the probable actions and traj of every other cars/peds/actors on the road.

So tesla is literally just getting around what some others have been doing for 5 years.

Now about automatic video labeling. Its has become the standard for a-couple years, you have to ask yourself why is Tesla only implementing it now?

 

[Bladerskb]

A good search engine is just natural language processing and a database. Anyone can do that, and there have been hundreds of academic papers written about it.

But if you do it at scale, then you've got Google. The constituent parts aren't revolutionary, but putting them together and doing it at scale is where the innovation happens.

uhh that is so far from the truth. Architecture matters actually more than the data. Google didn't get to where they are because they were doing things at scale. Google got to where they are because they were the one inventing things. Look up GANS and Transformers.

NN architecture is why you have this realistic sounding AI trained on just 10 hours of audio.
Yet no one for years have yet to match the quality.

 

[jeremymc7]

So tesla is literally just getting around what some others have been doing for 5 years.

I honestly think they just don't want to throw the buckets of cash at self driving like companies like Alphabet and others are willing to.

 

[diplomat33]

I honestly think they just don't want to throw the buckets of cash at self driving like companies like Alphabet and others are willing to.

So you are saying that Tesla is not serious about solving FSD?

 

[jeremymc7]

So you are saying that Tesla is not serious about solving FSD?

Yes. No wait, I didn't say anything like that.

All I said was that Tesla wasn't throwing the absurdly high buckets of money like some others are doing. There is nothing in that statement about being serious about FSD, or wanting to get to L4/L5.

 

[DanCar]

Yes. No wait, I didn't say anything like that.

Tesla autopilot team - 250 people. Making money for the company.
Waymo - 1,500 people. Losing nearly a billion dollars each quarter.

Elon has indicated he plans to add people to team, especially in China.

 

[Bladerskb]

Still to this day, the traffic control feature and cut-in detection are proofs of concept for FSD.

You disagreed with one of my posts. Yet you basically repeated what was written in it yourself. That there is no prediction network in the car. All they have is the cut-in detection. Now they are finally adding prediction which everyone else have been doing for years and have been using RNN/LSTM and other temporal networks and architecture that take into account time. Leave it to Elon to hype it as "4D". If that's the case, Waymo have been "4D"ing since 2014.

It blows my mind that Tesla can release such a feature while Apple and Google still struggle with simple dictation (mostly Apple, lol).

That's because you haven't looked into what goes into voice recognition and voice synthesis neural networks and their past and current quality and accuracy rates.

There's nothing special or amazing about Tesla's traffic control feature if you have been following the ML community.
You can create a traffic light recognition network and get it up to 90% accuracy with relative easy.

The fleet is indispensable for FSD once you understand how they were able to develop the traffic control feature through fleet triggers and learning.

To follow up what i said above. For example, mobileye's EyeQ4 from Q4 2017 has traffic light and relevancy of traffic light networks and its not geo-fenced.

 

[jeremymc7]

Tesla autopilot team - 150 people. Making money for the company.
Waymo - 1,500 people. Losing nearly a billion dollars each quarter.

Elon has indicated he plans to add people to team, especially in China.

Research often isn't cheap or "profitable". Doesn't mean it's wasteful or not worth the effort.

Not on either side of the fence here.

Owner since Roadster days and bought two Tesla's in last six months including FSD option. So you can't say I'm a hater.

 

[powertoold]

That there is no prediction network in the car.

This is an oxymoron.