Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Why Tesla's fleet learning is a big deal

Saghost

Well-Known Member
Oct 9, 2013
8,224
7,088
Delaware
@Paddy3101 - I think you missed a wrinkle. Elaborating...

As the performance of the "L5 FSD candidate system" improves, insurance companies and regulators will likely be motivated to incentivize humans to "let the car do it" to remove the unpredictable humans from the roadways. I expect much of the tail of edge cases to be directly caused by the "cohabiting" human drivers. If/when the human drivers are removed, the self-driving car performance numbers will improve overnight. I suspect that can be proven statistically in test markets pretty easily.

In other words, if you can get Jimmy to let the car drive he becomes safer and the self-driving cars around him become safer.

Yup. That's when shadow mode gets reversed, and to get cheap insurance you have to let the car watch as you drive and intervene if you go to do something stupid.

We're certainly not close to something like that yet - if nothing else, 98% of the new cars on the road don't have the equipment, and we currently turn the fleet every twelve years or so - but if FSD is as mature as it looks, I suspect we'll see nearly mandatory shadow boss mode in my lifetime in at least some jurisdictions.
 

KD4MNI

Member
Oct 22, 2018
23
19
GA
The unpredictable humans in the other vehicles, perhaps. Not driving the Teslas. In fact, the FSD system is "reprogrammed" by studying cases where human drivers took control from AutoPilot (dropouts are automatically recorded and reported) and the underlying cause identified, then the dNN is retrained using a large collection of images harvested from actual data to help the system learn how to properly respond. Wash, rinse, repeat. The process is partially automated, and improvements are more exponential.than linear. Not everything can be "fixed" with training, there are heuristics (lines of logical code) as well, it's essentially a hybrid system, which is part of what makes it so extraordinary, using the best of both worlds to solve an incredibly complex, difficult problem. Only with lots of tuning and training will the system eventually approach the skill levels of a situationally-aware human driver (but with the near-instant, untiring reflexes of a machine).
 

Saghost

Well-Known Member
Oct 9, 2013
8,224
7,088
Delaware
Unless FSD comes free with the car, I can't see the average buyer paying thousands for it. The vast majority of people want to get into any given car at the lowest monthly payment possible.

If Tesla can prove say half the accident rate, insurance companies will offer much lower insurance rates for cars equipped with it, which will drive those bargain buyer consumers to see it as an investment.
 
  • Like
Reactions: JeffnReno

StealthP3D

Well-Known Member
Dec 12, 2018
9,894
81,305
Maple Falls, WA
Good luck with that interpretation. It will not be able to handle all situations , not even close, by end of year. It will not handle someone giving hand gestures (cameras are low res), will not handle partially occluded objects like stop sign blocked by tree growth, will not handle adversarial attacks like when people learn putting a little white sticker on a stop sign means Tesla misreads it, etc...

@1375mlm didn't say "feature complete" means it could handle all situations without intervention. The edge cases you mention would almost certainly not be handled as soon as the system is "feature complete" (which is why a human driver would still be a requirement). "Feature complete" simply means they have developed the software for all the major types of driving (freeway, city, country, roundabouts, stoplights, etc.) but it will still need to be trained to handle the edge cases.

Elon has been wrong so many times on FSD and even he admits his timelines are overly optimistic.

Musk does tend to be overly optimistic on the timeline for complex tasks. But those who doubt his capability to achieve those complex tasks have been made a fool of time and time again.
 
  • Love
Reactions: CarlK and JeffnReno
  • Like
Reactions: 1375mlm

wcorey

Active Member
Jul 11, 2018
1,657
788
Killingly, CT
Unless FSD comes free with the car, I can't see the average buyer paying thousands for it. The vast majority of people want to get into any given car at the lowest monthly payment possible.
Oh, they'll pay for as NOTHING is free. Whether it's an optional extra or added to the base price of the car. At some point that's how it will come but, be clear on one thing, it won't be free. What will happen is there will, at some point, start restricting highways to self driving cars only. Most cars come with AC...it's not free either, it's part of the price.
 
  • Like
Reactions: JeffnReno

wcorey

Active Member
Jul 11, 2018
1,657
788
Killingly, CT
If Tesla can prove say half the accident rate, insurance companies will offer much lower insurance rates for cars equipped with it, which will drive those bargain buyer consumers to see it as an investment.
That raises a good question. How are insurance costs distributed? Are the safest cars the ones with the lowest insurance? A P3D, P100D model S. A car that can do 155mph, 0-60 in under 3 secs OR is it actuarially determined based on claims and/or historic claims. So long as a drunk 30 yr old can go all Mad Max in it or that's the driver they attract. I don't know as I am not an actuary for insurance companies. My Model 3 AWD non-p is significantly more expensive than any other car's insurance I had. I've got a spotless driving record. I suspect it has a lot to do with being significantly more expensive than any other car I've ever owned. I suspect it is more what would it cost to repair and how much are law suit settlements. Thoughts?
 
  • Like
Reactions: JeffnReno

Mo City

Active Member
Jul 17, 2016
1,910
11,883
near Houston
Unless FSD comes free with the car, I can't see the average buyer paying thousands for it. The vast majority of people want to get into any given car at the lowest monthly payment possible.
I completely disagree.

If Tesla's FSD product actually does full self driving in 2-3 years, it will be a truly compelling product. Tesla will increase the price dramatically and people will rush to pay for it.

That said, I don't believe the current sensors are adequate for Level 4/5.
 
  • Like
Reactions: CarlK

wcorey

Active Member
Jul 11, 2018
1,657
788
Killingly, CT
I completely disagree.

If Tesla's FSD product actually does full self driving in 2-3 years, it will be a truly compelling product. Tesla will increase the price dramatically and people will rush to pay for it.

That said, I don't believe the current sensors are adequate for Level 4/5.
We'll know in 6 months. Allegedly, regulatory approval ready in 18 months. In 6 months we'll know how likely approval in 12 months is. Until end of 2019, it's all pure speculation.
 

diplomat33

Average guy who loves autonomous vehicles
Aug 3, 2017
8,379
11,683
Terre Haute, IN USA
That said, I don't believe the current sensors are adequate for Level 4/5.

Level 4 yes
Level 5 probably not

Remember that L4 is highly automated where the car self-drive with no driver at all but only in certain situations. So Tesla could do L4 highway for example.

L5 is fully automated where the car can self-drive with no driver at all in all circumstances.
 

wcorey

Active Member
Jul 11, 2018
1,657
788
Killingly, CT
@1375mlm didn't say "feature complete" means it could handle all situations without intervention. The edge cases you mention would almost certainly not be handled as soon as the system is "feature complete" (which is why a human driver would still be a requirement). "Feature complete" simply means they have developed the software for all the major types of driving (freeway, city, country, roundabouts, stoplights, etc.) but it will still need to be trained to handle the edge cases.



Musk does tend to be overly optimistic on the timeline for complex tasks. But those who doubt his capability to achieve those complex tasks have been made a fool of time and time again.
There will always be edge cases, always. There are right now today driving a Kia or Audi. Any event you could and did not anticipate is an edge case. As for Elon bring over-optimistic, the closer you are to the finish line, the less wiggle room you have. He was unambiguous about this year. He equivocated slightly on end of 2020.
 

wcorey

Active Member
Jul 11, 2018
1,657
788
Killingly, CT
Level 4 yes
Level 5 probably not

Remember that L4 is highly automated where the car self-drive with no driver at all but only in certain situations. So Tesla could do L4 highway for example.

L5 is fully automated where the car can self-drive with no driver at all in all circumstances.
I'd put money on yes to both by end of 2020, soft yes by end of 2019. BTW, how do you find the new 2019-12-1?
 

diplomat33

Average guy who loves autonomous vehicles
Aug 3, 2017
8,379
11,683
Terre Haute, IN USA
I'd put money on yes to both. BTW, how do you find the new 2029-12-1?

I hope Tesla does L5. I just know that it is more difficult because it has to do self-driving in all circumstances.

Regarding thew new update, overall, I really like it. I don't care for the atari games and I probably won't see much difference with the peak torque performance. Bu I really like being able to see if you have an update pending in advance. I also really like the increase supercharging and battery warm up. That should make long trips even better.

Autopilot does seem a bit better too. TACC is super smooth in normal traffic. Auto steer is solid. Maybe this isn't new, but I did notice that on local streets, auto steer seemed to lock on to the center divider line so it stayed straight and did not veer into parked cars. There was one instance on a city street where there was no right lane line at all (it was just grass and driveways), left blue line was solid but the right blue line on the screen was undulating a bit and AP asked me to take control but the car still stayed perfectly straight though and never had to disengage AP, I just tugged the wheel. I also noticed coming into the entrance of my work which is a wide entrance, wide enough for 2 cars but no center lane, that when I hit the turn signal to make a turn, AP warned me with the red line on the screen. So even though there was no center line, AP still interpreted it as a lane change and since there were cars coming my way, it warned me against a "lane change". Neat!

I do get a chance to try NOA later today on the new update, when I go see Endgame. :)
 
Last edited:
  • Like
Reactions: JeffnReno

wcorey

Active Member
Jul 11, 2018
1,657
788
Killingly, CT
I hope Tesla does L5. I just know that it is more difficult because it has to do self-driving in all circumstances.

Regarding thew new update, overall, I really like it. I don't care for the atari games and I probably won't see much difference with the peak torque performance. Bu I really like being able to see if you have an update pending in advance. I also really like the increase supercharging and battery warm up. That should make long trips even better.

Autopilot does seem a bit better too. TACC is super smooth in normal traffic. Auto steer is solid. Maybe this isn't new, but I did notice that on local streets, auto steer seemed to lock on to the center divider line so it stayed straight and did not veer into parked cars. There was one instance on a city street where there was no right lane line at all (it was just grass and driveways), left blue line was solid but the right blue line on the screen was undulating a bit and AP asked me to take control but the car still stayed perfectly straight though and never had to disengage AP, I just tugged the wheel. I also noticed coming into the entrance of my work which is a wide entrance, wide enough for 2 cars but no center lane, that when I hit the turn signal to make a turn, AP warned me with the red line on the screen. So even though there was no center line, AP still interpreted it as a lane change and since there were cars coming my way, it warned me against a "lane change". Neat!

I do get a chance to try NOA later today on the new update, when I go see Endgame. :)
On L5, that is all circumstances a normal driver would be expected to successfully drive. No raging blizzards, etc. I didn't even like Atari back when it was 'a thing'. I was making IMDOS (Imsai's microcomputer OS) reentrant. Actually that was earlier than Atari. I hope I get 2019-12-1 soon, I have several longish trips this week. One of which is to see Endgame, on Monday, late the day 1 crowds have day 1 and family day Sat/Sun.

Of my last several posts, here's my philosophy, which I came up with over my career in software. It basically boils down to problem decomposition. Any problem space can be broken down into a series of knowns with interspersed unknowns. The more unknowns there are the more volatile the estimates are. The bottom line is as you approach completion the unknowns approach zero and the estimate becomes more reliable. The way Musk addressed where we'd be end of this year there was no equivocation. Short of OCR for speed limit, yield, stop signs (a known) NoA is essentially good for local driving. I use autosteer all the time. The only thing the car doesn't do is stop at intersections. TACC really should, upon seeing a new geocoded speed limit, set current speed to that...it's a tad flaky in my opinion. So, yeah, from a software point of view, I feel comfortable they know how to do what's remaining. So the problem becomes essentially an assignment matrix aka optimize the resource assignments to minimize job completion.
As for being able to cap steering wheel columns by end of next year... that is dependent on things outside their control. For instance. the town improperly geocoded our house location so the car, rightfully, is expecting our driveway where it isn't located. That's not a Tesla problem, that's a local government problem. Properly opening and closing garage doors is not a Tesla problem, it's a HomeLink problem.
 
  • Like
Reactions: diplomat33

wcorey

Active Member
Jul 11, 2018
1,657
788
Killingly, CT
Oh, they'll pay for as NOTHING is free. Whether it's an optional extra or added to the base price of the car. At some point that's how it will come but, be clear on one thing, it won't be free. What will happen is there will, at some point, start restricting highways to self driving cars only. Most cars come with AC...it's not free either, it's part of the price.
I want to elaborate on this. Simply put, software is difficult, pricing wise. Hardware, easy you know how much the raw materials cost, you know your chosen margin and, if marketing is doing their job, they know initial customer base, maturation curve and aging fall off. Software has N resources assigned, a competency curve where the problem space goes from heavily unknown next steps to few to zero unknowns. Then there is "the march of 9s", processing new data, extracting new edge cases etc. It's very difficult to come up with an end of lifecycle date. If they are shooting for 5 nines, carrier grade, or better? The only way it will get there is to eliminate human drivers. Here's the problem there, better than the avg driver, who's avg? I've been driving over 50 years. There are 'edge cases' I've never encountered. I've encountered more than the avg driver though. Answering that is a societal problem. There's a chance I could find myself in a fatal accident. It's small, way smaller than for a 16 year old but greater than zero. If I am under FSD and get into an accident someone doesn't live through, is that an actuarial (statistical) problem or a fault with the software/hardware? As Elon posited, if you address all the edge cases then we are living in a simulation. As it relates to the ongoing cost of FSD, essentially, who defines 'done'. In the mean time, you have a staff and assets working on that that isn't free. The cost needs to be past on.
 
Last edited:

StealthP3D

Well-Known Member
Dec 12, 2018
9,894
81,305
Maple Falls, WA

StealthP3D

Well-Known Member
Dec 12, 2018
9,894
81,305
Maple Falls, WA
worked fine for me.

That's odd. I've been trying for two days to open it with Microsoft browsers (Edge and IE 10) and it doesn't work. It opened right away when I used Firefox.

Good story. But it looks like Tesla is at least a couple of years ahead of the most advanced machine vision/AI researchers in academia!
 

HAWX 12

Member
Jan 3, 2019
75
39
Kailua Hawaii
Is my car "always" giving feedback when it experiences a situation it can't handle (on autopilot)? Or do i have to bug report each time it panics and kicks off?
I'm hoping I am contributing to the data collection.
 

Products we're discussing on TMC...

About Us

Formed in 2006, Tesla Motors Club (TMC) was the first independent online Tesla community. Today it remains the largest and most dynamic community of Tesla enthusiasts. Learn more.

Do you value your experience at TMC? Consider becoming a Supporting Member of Tesla Motors Club. As a thank you for your contribution, you'll get nearly no ads in the Community and Groups sections. Additional perks are available depending on the level of contribution. Please visit the Account Upgrades page for more details.


SUPPORT TMC
Top