Tesla autopilot HW3

[mongo]

High degree? That video is filled with errors and false positives. Plus its missing traffic lights, traffic signs, road markings, road signs, stop line, intersection markings, road profile, animals, hazard, general object detection, etc. Which are all essential to driving, insinuating that the current NN solved vision is ridiculous.

They are not insinuating that the current Tesla NN solved vision at all. Quite the opposite, as stated the the very text you quoted.:

But I wouldn’t say this proves Tesla has solved vision, since the devil is in the 0.01% of failure cases, which can’t be seen from a short video.

 

[Cirrus MS100D]

No this really is a firmware version. We’ve seen ones with dev/ prefaces too that don’t have versions or git hashes. Someone previously explained that “terminal” is the last station before the car gets delivered, so it’s basically like factory firmware used pre-delivery.

Doesn’t the TeslaFi owner/operator lurk on the forums? Wouldn’t they be able to tell us (or “Not able to disclose” tell us) what happened here?

 

[Bladerskb]

They are not insinuating that the current Tesla NN solved vision at all. Quite the opposite, as stated the the very text you quoted.:

That statement to me reads as though getting to 0.01% was tesla's main problem at the time of the video when its actually adding those detectors i listed.

The point I was trying to make is that, assuming Tesla's cameras and radar are sufficient, then the worst case roll out for FSD, when it exists, would be an AP computer swap. If HW3 is sufficient, then it is only an OTA update for current 3 production.

"When" is exactly the problem and has been since 2015. Let me give you a hint, it will never happen in any timeframe that elon gives until it eventually happens. Why? Because elon doesn't speak for the AP team, he makes his prediction for personal gain. Remember when he said they will only need 10 million miles of "validation" and then they will disable confirmation. He made it sound like something trivial that didn't require any additional engineering or development and that it was imminent. Here we are 5+ months later and nothing. People have to start understand that the AP have their own delivery schedule. Since they are agile they probably do two week springs. They already know what they are going to do 6 months in advance. Elon doesn't make prediction off these internal timelines. He makes predictions that will make him look good. Kinda like Trump. That's why as you have said he will never say more than "2 years" because it won't make him look good. If he said and kept saying 6 years in 2015, he will sound like everyone else. It doesn't make him look like the best and ahead of everyone else.

For anyone else, the cars do not currently have HW on them. So, once a system is proven on test vehicles, that hardware needs to be integrated into the vehicle, manufacturing process, and supply chain. Likely aligned to a model year change. So that is a longer delay after they solve FSD than what Tesla may have.

True they don't have the HW in cars right now but alot of people have a set date for when their self driving platform will be ready. They are building their autonomous system on completely new platform that supports SAE level 3-5. The timeline for that is 2020-2021 for the companies i listed. Once they have the platform, there is no longer 3-4 years lead time for a new model year. Every model that uses the platform will have access to the platform. Also when the HW comes in cars in 2021, they will also come with the software so there won't be a waiting period. I can't speak for everyone else but i can speak for Mobileye close partners.

 

[mongo]

Let me give you a hint, it will never happen in any timeframe that elon gives until it eventually happens. Why?

Cause that is the way software goes in general

 

[Bladerskb]

Cause that is the way software goes in general

I'm a senior software engineer that have worked for several companies, having used both waterfall methodology and currently now using agile, doing two week scrum sprints. I know first hand what realistic timelines,unrealistic timelines and what straight up lying for PR reasons timelines are.

 

[mongo]

I'm a senior software engineer that have worked for several companies, having used both waterfall methodology and currently now using agile, doing two week scrum sprints. I know first hand what realistic timelines,unrealistic timelines and what straight up lying for PR reasons timelines are.

Fun, I'm a multi-disciplinary consultant who has dealt with fixing problems in other people's code.

Genuine industry question:
How do you set realistic project timelines for things you have never done before? 2 week sprints I get, but on a multi-year project, do you just plan to increase hours to hit the intermediate targets, or what?
Even short term projects can get thrown off when a dependency is broken.

 

[HolyGrail]

@Bladerskb what is your area of expertise?

 

[electronblue]

Fun, I'm a multi-disciplinary consultant who has dealt with fixing problems in other people's code.

Genuine industry question:
How do you set realistic project timelines for things you have never done before? 2 week sprints I get, but on a multi-year project, do you just plan to increase hours to hit the intermediate targets, or what?
Even short term projects can get thrown off when a dependency is broken.

In your heart of hearts, do you really think Elon simply made a bad estimate (and has done so time and again) for the software project that is EAP/FSD? Or could there have been other reasons why his estimates turned out to be bad.

Estimating software timelines is hard for sure. But it is also possible for people to come up with timelines for reasons completely unrelated to difficulty of estimating them. And none of this explains why Elon never revisits these continuous timeline mistakes and clarifies or explains them.

 

[diplomat33]

Tesla has just paused orders for all models (S,X and 3) and teasing a big announcement today at 2PM (PT). The pause in orders for all cars would suggest that Tesla is making a change to all cars. I am thinking we are going to get an announcement about AP3.

 

[electronblue]

I don’t think there is a proven path forward for conventional, hand-coded software solving complex robotics problems involving human interaction. There are some proofs of concept for imitation learning and deep reinforcement learning. But not for conventional software.

That’s why I see more promise in Tesla’s production fleet learning approach than in Waymo’s small-scale testing approach. Waymo can’t collect enough data on the state-action pairs of human driving to do imitation learning across for all driving subtasks. If anyone can collect enough data, it’s Tesla.

If you think driving is a machine learning problem rather than a conventional software problem, Tesla has the most promising approach. Machine learning requires large datasets; Tesla has access to them, and Waymo doesn’t.

That is all fine and well if you assume Tesla is in a position to collect enough relevant data — and that the data they can collect is relevant...

However the picture changes if one assumes the data realistically gathered from consumer cars is not really that relevant for the purpose you are outlining here ie training NNs. If so, suddenly for the thesis it means Tesla does not have a data advantage in training its NNs and at the same time they are also behind in other aspects...

Same if the amount of data turns out not to be relevant. If either of these assumptions fail, turns out Tesla is merely behind. And those assumptions are far from clear in many minds... to put it kindly.

That said I do acknowledge Tesla has a deployment and validation advantage to a degree. But the idea that they have a NN training advantage is far from agreed upon.

 

[iTech]

Tesla moving Smart Summon, Autopark, and Navigate on Autopilot to FSD.

I can already hear the lawyers salivating.

 

[Kant.Ing]

All because these features use more cameras. Or wait for the actual HW3.

 

[DDotJ]

Tesla moving Smart Summon, Autopark, and Navigate on Autopilot to FSD.

I can already hear the lawyers salivating.

@Elon on Twitter: Everyone who bought Enhanced Autopilot will get Navigate on Autopilot on highways, advanced summon & autopark

 

[diplomat33]

@Elon on Twitter: Everyone who bought Enhanced Autopilot will get Navigate on Autopilot on highways, advanced summon & autopark

Good to know. I do wonder if it is worth buying FSD now since the only new features will be the city driving with traffic lights. Of course, buying FSD now will also get us the AP3 chip which may improve NOA, Advanced Summon and Auto Park. Plus, Getting FSD will get us the future updates that improve city driving.

 

[mongo]

In your heart of hearts, do you really think Elon simply made a bad estimate (and has done so time and again) for the software project that is EAP/FSD? Or could there have been other reasons why his estimates turned out to be bad.

I think he did not intend to provide bad estimates. Whether due to ignorance on his part, or over-optomistic assumptions about progress/ team effort, he really thought it was possible in 2ish years.

I say that because I could have done the same thing, and have on smaller scales. (On 4th year of building a pole barn)...

 

[rnortman]

I think he did not intend to provide bad estimates. Whether due to ignorance on his part, or over-optomistic assumptions about progress/ team effort, he really thought it was possible in 2ish years.

I say that because I could have done the same thing, and have on smaller scales. (On 4th year of building a pole barn)...

The problem is that the demo video was clearly a lie. It was not implying that there was any software development let to do, only validation and regulatory approval. Elon may have been optimistic about the schedule, but he must have known without a doubt that this was a lie.

 

[electronblue]

I think the chances exist within the realm of reason that Elon knew he was giving out a bad estimate. We do not know of course but I can not really with any certainty chalk this up to ”software estimation is hard” either. I guess the best we can do is say we don’t know.

We do know there are reports the Autopilot team was not happy with Elon’s announcement and I would be surprised if nobody had told him beforehand...

 

[mongo]

The problem is that the demo video was clearly a lie. It was not implying that there was any software development let to do, only validation and regulatory approval. Elon may have been optimistic about the schedule, but he must have known without a doubt that this was a lie.

My take was that the video showed that the hardware could do self driving. This was when AP1 was discontinued and the feature set was reduced as they call out in the blog post that accompanied the video:
All Tesla Cars Being Produced Now Have Full Self-Driving Hardware

We are excited to announce that, as of today, all Tesla vehicles produced in our factory – including Model 3 – will have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver.

Before activating the features enabled by the new hardware, we will further calibrate the system using millions of miles of real-world driving to ensure significant improvements to safety and convenience. While this is occurring, Teslas with new hardware will temporarily lack certain features currently available on Teslas with first-generation Autopilot hardware, including some standard safety features such as automatic emergency braking, collision warning, lane holding and active cruise control. As these features are robustly validated we will enable them over the air, together with a rapidly expanding set of entirely new features.

 

[mongo]

We do know there are reports the Autopilot team was not happy with Elon’s announcement and I would be surprised if nobody had told him beforehand...

One common theme in all of Elon's companies is that people never like his timelines. He always pushes dates forward (see also the recent SpaceX firings). Hes a 18+ hr/day kind of programmer and expects that out of everyone else....

 

[electronblue]

One common theme in all of Elon's companies is that people never like his timelines. He always pushes dates forward (see also the recent SpaceX firings). Hes a 18+ hr/day kind of programmer and expects that out of everyone else....

All good and well until you sell that promise to someone.