FSD v12.x (end to end AI)

[Matias]

You are asking me to do your home work

Miles between incidents for regular cars (according to Cruise’s modified version of Virginia Tech study, which removes crashes on highways and rural roads):​

​

Miles between Level 1 incidents: 282,485​

​

Miles between Level 1 and 2 incidents: 99,404​

​

Miles between Level 1, 2 and 3 incidents: 20,661​

​

Miles between Level 1, 2, 3 and 4 incidents: 9,21(8)​

​

​

https://www.reddit.com/r/SelfDrivingCarsLie/comments/bxx7ea/cruises_secret_apollo_robotaxi_milestones_paywall

Yes I saw that reddit post, but I don’t consider it as a reliable source
Unless there is some link behind the paywall.

 

[EVNow]

Yes I saw that reddit post, but I don’t consider it as a reliable source
Unless there is some link behind the paywall.

Thats because you didn't click the link in the Reddit post that takes you to a paywalled Information article. Redditor reposted the entire article on reddit ....

 

[Ben W]

Doesn’t suck for me. I look at V12 as a significant upgrade. But not the best thing since sliced bread nor cold fusion level breakthrough.

V10 - 1 disengagement in 5 miles - 2022
V11 - 1 disengagement in 10 miles - 2023
V12 - 1 disengagement in 20 miles. - 2024

So, when will we get to RT level and how ? Both V10 and V11 quickly hit a plateau and needed major architectural changes to break out of.

My guess is that v12, being E2E trained, will have quite a lot more headroom than the earlier brittle heuristic-based approaches. Perhaps the current approach will eventually reach 1 disengagement in 1000 miles. But I don't think it can plausibly reach RT level (1 safety-critical disengagement/failure per million miles). The limiting factors are:

- Robustness to poorer environmental conditions (rain, dirt, fog, sun glare, very low light, etc). Radar/lidar may help, or higher dynamic range cameras, or more/redundant cameras, or laser glass cleaning.
- Compute. Can the HW3/HW4 computers hold and process large enough networks to contain all the necessary knowledge for all the edge cases that the car doesn't yet know how to handle? It would be really nice if HW3/HW4 cars would eventually have an upgrade path to a faster computer, but I'm not holding my breath.
- Latency. (Related to compute.) With larger and more powerful networks comes slower processing and reaction time. Adding radar/lidar for instant 3D mapping, or a parallel faster "reflex" network to more quickly process evasive maneuvers and collision avoidance/minimization, may help a lot with this.
- Time horizon. By training only on very short clips/scenarios, the NN is trained to be instinctual/reactive, but has no way to gain longer-term "general intelligence", which is required to properly solve a surprising number of real-world driving problems, in particular to gain an intuitive understanding of other drivers' intentions and goals.

I'll be very interested to see what HW5 consists of. My hope (and perhaps Elon's secret hope) is that NHTSA may mandate Lidar/radar in some form for L3/L4 operations, which would vastly improve Tesla's hardware suite, without making Elon lose face.

 

[Ben W]

Collect the data as the potholes are hit, phone home, and share that with other cars. Same for when the potholes are filled. Tesla could do so much with vehcle data collection, and I know lots of us are confused as to why they don't do it. Perhaps too much LTE data would become prohibitively expensive.

Now we know why Starlink is adding cellular coverage. Well, maybe not.

Reporting a pothole is a trivial amount of information; it would fit in a text message. (Just the GPS coordinates, and perhaps a few bytes to describe the size / position of the pothole in the lanes.) The car should be able to do much of the real-time detection and avoidance on its own, but with a pothole database it could slow down in advance of reaching the spot, both to better avoid the pothole if it's still there, and also to effectively gather more data and detect if e.g. the pothole has been fixed. This approach could work for all manner of road debris or mapping exceptions (e.g. temporary signs).

 

[Ben W]

A pilot friend shared one with me that I now cannot get out of my head any time I hear someone misuse the word: when the flight attendant announces "we'll be landing momentarily", I picture the pilot doing a touch-and-go...

To be fair, "imminently" is also one of the accepted definitions: Definition of MOMENTARILY

Obligatory viewing:

 

[EVNow]

Reporting a pothole is an inconsequential amount of information; it would fit in a text message. (Just the GPS coordinates, and perhaps a few bytes to describe the size / position of the pothole in the lanes.) The car should be able to do much of the real-time detection and avoidance on its own, but with a pothole database it could slow down in advance of reaching the spot, both to better avoid the pothole if it's still there, and also to effectively gather more data and detect if e.g. the pothole has been fixed. This approach could work for all manner of road debris or mapping exceptions (e.g. temporary signs).

Waze already does some of this. Tesla has to just copy the approach.

 

[Daniel in SD]

Yes.

View attachment 1046461

Autonomous Car Progress

I think supervised FSD is a terrible idea, and is going to fall on its face so I agree with you on that one. I don't doubt that some version of FSD (or a competitors version) will be safer than the average human driver, but I don't think that's a good comparison. The entire comparison to...

teslamotorsclub.com

Not very useful for robotaxi safety because the distribution might be completely different (it is for Waymo).

 

[EVNow]

Not very useful for robotaxi safety because the distribution might be completely different (it is for Waymo).

How so ?

 

[Daniel in SD]

How so ?

Robotaxis will not have any L4 collisions so they’re not relevant.

 

[EVNow]

Robotaxis will not have any L4 collisions so they’re not relevant.

Yes - we saw how Cruise manipulated the stats by "getting stuck" instead of proceeding when it wasn't sure there won't be a collision.

 

[sleepydoc]

Biggest difference for me is the car doesn’t slam the brake when someone pulls out in front of me a 1/4 mile up the road

That’s been improving for a while but IME v12 does a near-perfect job of slowing enough to miss the car without slowing excessively.

 

[sleepydoc]

There’s no incentive for it.

AP/NOA transitioning to v11 first is more likely.

That keeps the Tesla FSD mileage curve looking good.

While v12 may ultimately be better on the highway, Tesla will need those miles from v11 FSD before v12 is validated sufficiently.

Around here we have a lot of cloverleafs with short merging sections that require you to plan your merge as you approach and coordinate slowing and accelerating traffic. (Yes, it’s a crappy road design, but it’s what we got.) v11 just can’t handle these. It waits for the dotted line before it starts and by the time it’s done thinking about merging it’s too late. V12 would be a great improvement in these situations.

 

[aronth5]

You are asking me to do your home work

Miles between incidents for regular cars (according to Cruise’s modified version of Virginia Tech study, which removes crashes on highways and rural roads):​

​

Miles between Level 1 incidents: 282,485​

Miles between Level 1 and 2 incidents: 99,404​

Miles between Level 1, 2 and 3 incidents: 20,661​

Miles between Level 1, 2, 3 and 4 incidents: 9,21(8)​

​

I guess I'm doing pretty well. 845k miles driven. Zero accidents of any kind. My insurance company has made a lot of money off of me.

 

[Singuy]

Not the point. Do you acknowledge that others are having these issues ?

Can we get this on video with multiple cars doing this at the same location? How reproduceable is this?

When people talked about lane wiggle while getting into turning lanes...I have experienced this
When people talked about "lane drift" on v11, I have experienced this

However this not keeping lane thing while going straight is something I and none of the people I talk to with FSD have experienced. I mean sure it can happen, but is it a car issue? Or maybe it's a particular kind of lane? Maybe it's dodging something at the time but you didn't notice? It's a mystery for many of us.

 

[joelkfla]

Doesn’t suck for me. I look at V12 as a significant upgrade. But not the best thing since sliced bread nor cold fusion level breakthrough.

V10 - 1 disengagement in 5 miles - 2022
V11 - 1 disengagement in 10 miles - 2023
V12 - 1 disengagement in 20 miles. - 2024

So, when will we get to RT level and how ? Both V10 and V11 quickly hit a plateau and needed major architectural changes to break out of.

You must not be making many ULT's into divided highways.

 

[joelkfla]

I had a similar experience the other day when a car wanted to pull out from a CVS parking across 2 lanes to turn left. FSD stopped short of the line of cars in front of me and let the car cross to make the turn. At first I wasn't sure why FSD stopped until I saw the car trying to come out. I was shocked but in a good way.

V12 also consistently avoids pulling into intersections at lights to avoid blocking traffic in case the light turns red.

This seems courteous, but is actually unsafe behavior if there is a lane of traffic to your left. The driver exiting the driveway may feel pressured to pull out, even though he can't see whether cars are approaching in the left lane (or a 2-way left turn lane.)

 

[APotatoGod]

And while we are at it
They're, Their, There.
effect, affect
then, than
...

There, there. Don't let their affect affect you.

 

[hey8you]

Tried V12.3.6 today when I went to grab lunch. It tried to do a lane change too close to some road pylons(or whatever they are called). If I hadn't intervened, I think it would have hit them. It tried to do it twice and both times I intervened. There also was some hesitation when trying to enter a turn lane. It seem to have trouble picking a lane so I did intervene. The rest of the trip was no issue. I did get to try a U-turn. It handled it great.

 

[EVNow]

Can we get this on video with multiple cars doing this at the same location? How reproduceable is this?

I think almost all of us have had this experience. So, you are the outlier here claiming perfect drives for weeks.

It is not a “lane keeping” issue in the sense of Ford blue etc. This is where intentionally FSD is driving to one side of the lane (for whatever reason) and sometimes going over the edge. For eg., yesterday FSD was driving almost on the right line of the two lane road because there was a F150 driving on the left lane. In general I’d only go that far away if it’s a huge truck, not F150. So, FSD needs more specific training to learn this offset rule like humans do.

 

[AlanSubie4Life]

Around here we have a lot of cloverleafs with short merging sections that require you to plan your merge as you approach and coordinate slowing and accelerating traffic. (Yes, it’s a crappy road design, but it’s what we got.) v11 just can’t handle these. It waits for the dotted line before it starts and by the time it’s done thinking about merging it’s too late. V12 would be a great improvement in these situations.

Seems likely, but as I said, just no incentive.

Obviously going to take a year or so before they're ready for that and they'll need the valuable FSD miles for their graphs well before that. Best possible function is not the priority!!!