Tesla autopilot HW3

[EVNow]

I’m sure there are customer tables to join in as well.

I'm talking about "vehicles". It should be "vehicle". So, the table name should be Person, not People, for eg.

This is all a very long-winded way of saying that the time to consider “cost and reliability” has long since passed, since that’s something that you do before you start rolling it out to customers.

BS.

I'm talking about scheduling in such a way as to reduce cost. For eg. now they are apparently upgrading when someone is already at the shop for a different reason and servicing already exposes the computer.

When they start having the Mobile tech do the upgrades, for eg., it would make much more sense to upgrade people who live close by rather than by the date when they bought the car.

 

[EVNow]

And unlike the lane merger problem that can be easily managed by taking control, how will failure to stop at a stoplight or blowing through a stop sign or yield sign work out?

We'll have to monitor this closely. If the car is not slowing down well before a stop sign or traffic light (or not showing those in the visualization), time to take over.

 

[GlmnAlyAirCar]

and rightly so. That's what we were promised, after all.

Exactly

 

[azred]

We'll have to monitor this closely. If the car is not slowing down well before a stop sign or traffic light (or not showing those in the visualization), time to take over.

In my opinion, my two cars already brake very late when stopping behind cars and I am a pretty aggressive driver. (My wife would substitute “very” for “pretty.”) If that late braking normal behavior continues on FSD when reading a stoplight, I have a hunch I will be taking control on a regular basis, which makes FSD a joke.

 

[diplomat33]

While I have two Teslas needing the HW3 upgrade, I am not sure why anyone is that worried about getting the upgrade anytime soon.. Does anyone seriously believe Tesla will be able to flawlessly handle stoplights and various signage anytime in the next two years? Heck, my cars still go crazy when two lanes merge and that is surely child’s play compared to handling stoplights and signage. And unlike the lane merger problem that can be easily managed by taking control, how will failure to stop at a stoplight or blowing through a stop sign or yield sign work out? I use autopilot a lot and view it as a 15 year old beginning driver. It has great vision and reflexes, but requires adult supervision. The various quirks on freeways happen slowly enough to manage as long as I am paying attention but that all changes with FSD on the streets.

Sure, I believe Tesla will have flawless stop light and signage in 2 years. They already have a working version in their AP3 development cars now. I think another 2 years is enough to perfect it.

Keep in mind that the current Autopilot in our cars is running on AP2 (or AP2.5) which does not have the computational power for FSD. So it is not a good indicator of what AP3 will be capable of.

 

[electronblue]

Sure, I believe Tesla will have flawless stop light and signage in 2 years.

They need to have it a lot sooner for those robotaxis without anyone in them in 2020...

 

[timewasted]

I'm talking about scheduling in such a way as to reduce cost. For eg. now they are apparently upgrading when someone is already at the shop for a different reason and servicing already exposes the computer.

When they start having the Mobile tech do the upgrades, for eg., it would make much more sense to upgrade people who live close by rather than by the date when they bought the car.

Thankfully for me I’m perfectly happy dropping my car off at the service center in the morning and picking it up in the evening (or whatever time frame they prefer). My reason for being there is “my full self driving doesn’t seem to be working”.

I also pre-paid $3000 for this two years ago, so they’re not going to lose money on my appointment!

 

[diplomat33]

They need to have it a lot sooner for those robotaxis without anyone in them in 2020...

True. Which is why I believe Tesla will get those features done much sooner than 2 years.

 

[CK_Stuggi]

The mobileye patent specifically mentioned camera, vision recognition through training a neutral net IIRC. (I'll gladly accept correction on this).
Currently speeds for AP2+ are only map data based. It may be the data is somewhat accurate in large parts of the US, but for me in Germany the map data is outdated even for decade-old road sections. And I'm in an area where Daimler and Porsche are doing their testing, you bet there's up-to-date data available.
My 2015 AP1 did a pretty good job, definitely better than my 2017 AP2.
As regarding installation of AP3 hardware, European AP2 owners will be the absolute last to receive the upgrade as regulation is even stifling current possible functionality. So the argument "software which requires HW3" will likely apply there last. I'd venture to say it'll only become available over here once most AP2 cars are out of warranty.

 

[EVNow]

In my opinion, my two cars already brake very late when stopping behind cars and I am a pretty aggressive driver. (My wife would substitute “very” for “pretty.”) If that late braking normal behavior continues on FSD when reading a stoplight, I have a hunch I will be taking control on a regular basis, which makes FSD a joke.

Yes, the cars stop late for my taste as well - but I know when to expect the car to start braking, so I'm ready if it doesn't slow down. Basically if you have been using TACC, you can use the City NOA.

Infact, City NOA would be better. Visualization will show us when the car recognizes a stop sign or traffic light. No visual would be a trigger for braking.

 

[conman]

Never, fear, it will make up for the late braking for real red lights by lots of early phantom braking for fake red lights.

 

[DrDabbles]

AP1 has a processing power of a potato and it does it.

AP1 does not work like AP2+ at all. So it needs significantly less "processing power" to get its job done. Comparing the two in this way really doesn't make sense.

 

[DrDabbles]

Yeah, I totally understand that about Texas. I’ve been following closely. What I am saying is that the issues I had with customer service didn’t have to do with what you mentioned.

Overall, the system is overloaded. On the customer end you feel it more especially where I live as it is covered by the broader Texas mobile service division. Poor communication, had appt cancelled twice less than two days before the appt as correct parts weren’t ordered ect. It’s all stuff that isn’t specific to Texas but felt more as there is no phone contact for my service lite center and email communication is sparse. All growing pains that will improve.

I agree that service is spread thin. But the point I was trying to make is that with Texas being so hostile toward Tesla for so long, why would Tesla have invested in a service network only to risk it being outlawed shortly after?

Since we don't know what the spread of their customers is physically in the state, we can't really say for certain whether they're messing up or not. They're in the major areas you'd expect- Austin, SAT, Corpus Christie, Houston, Dallas, etc. But for those customers in Texas that are wealthy enough to own a Tesla and choose to live a billion miles from a city, there isn't really a good answer and there probably will never be. Except to stop wasting resources by living so far from civilization.

Anyway, hopefully Tesla service in general gets better constantly, and they take customer complaints about service seriously.

 

[electronblue]

True. Which is why I believe Tesla will get those features done much sooner than 2 years.

You may be disappointed.

 

[electronblue]

AP1 does not work like AP2+ at all. So it needs significantly less "processing power" to get its job done. Comparing the two in this way really doesn't make sense.

There is certain logic to the point nervetheless: MobilEye (AP1’s EyeQ3 but also later EyeQ4 etc included) has figured out a very power and energy efficient way of doing what Tesla Vision seems to be throwing major computing power at instead. So some genuine differences in approach too.

Of course from a Tesla computer perspective AP1 beyond EyeQ3 uses minimal processing power (in the MCU?) but EyeQ3 does the heavy lifting with lot less than what AP2+ needs for similar image recognition (or worse).

 

[dhanson865]

While I have two Teslas needing the HW3 upgrade, I am not sure why anyone is that worried about getting the upgrade anytime soon.. Does anyone seriously believe Tesla will be able to flawlessly handle stoplights and various signage anytime in the next two years? Heck, my cars still go crazy when two lanes merge and that is surely child’s play compared to handling stoplights and signage. And unlike the lane merger problem that can be easily managed by taking control, how will failure to stop at a stoplight or blowing through a stop sign or yield sign work out? I use autopilot a lot and view it as a 15 year old beginning driver. It has great vision and reflexes, but requires adult supervision. The various quirks on freeways happen slowly enough to manage as long as I am paying attention but that all changes with FSD on the streets.

10x the processing power in the HW3 board vs HW 2.x. If you think they can't improve the software with 10x the processing power you don't understand computing.

 

[GlmnAlyAirCar]

10x the processing power in the HW3 board vs HW 2.x. If you think they can't improve the software with 10x the processing power you don't understand computing.

Unless it’s running identical software at the same sample rates. Then it’s just unused processing power

 

[DrDabbles]

There is certain logic to the point nervetheless:

You can absolutely compare the features of the system, and you can compare energy consumption too. But to compare the processing necessary to operate one versus the other doesn't really work because of how vastly different they are implemented.

MobilEye (AP1’s EyeQ3 but also later EyeQ4 etc included) has figured out a very power and energy efficient way of doing what Tesla Vision seems to be throwing major computing power at instead. So some genuine differences in approach too.

This is where I'm really shining the light. EyeQ3 and EyeQ4 do not do what Tesla's system does. The end result is somewhat similar in that they both offer apparently evenly matched driver assistance products, but appearances are deceiving. Remember that EyeQ did not offer all of the features that Tesla wanted, so they added their own features on top of MobilEye's line. And again, they function on such a fundamentally different approach that they aren't really comparable. Saying "some genuine differences in approach too" misses that fact entirely. They could not be more dissimilar if you tried.

Of course from a Tesla computer perspective AP1 beyond EyeQ3 uses minimal processing power (in the MCU?) but EyeQ3 does the heavy lifting with lot less than what AP2+ needs for similar image recognition (or worse).

I think AP1 is taxing the MCU very heavily, and they've effectively stretched it to its limit. But to be clear here, "heavy lifting" in this case means line follow robot with ineffective radar object detection/avoidance that Tesla replaced, and speed limit sign recognition. That's not really all that impressive. In fact, you can do quite a lot of that "heavy lifting" with a raspberry pi or even an arduino if you'd like.

 

[dhanson865]

Unless it’s running identical software at the same sample rates. Then it’s just unused processing power

Have you never heard of a code fork? Parallel development?

If you have 10x the computing power you can't dev the new code on the old hardware. You start building that software using more capable hardware.

Yes the same code on old and new hardware won't change things. But once you have the new hardware you are eligible to run the new code base when it is released.

If they are bothering to do field upgrades you can bet they have code for the new hardware well along in the dev process.

 

[diplomat33]

You may be disappointed.

Or not. Stop being so pessimistic.