FSD Beta - Geographic Differences?

[BigRedDog]

I just got the Beta in the “100 Score” wave. I am located in the Midwest. It functions pretty poorly, and far differently than most of the videos I’ve been watching on YouTube.

I think Elon recently tweeted that functionality in locations outside of California would improve as more data from those locations was received.

Why would this matter? I thought FSD would supposed to be a “general solution” not reliant on HD map data. If that’s correct, why does geography matter? Roads in California are made of the same concrete, lane lines and stop lights as roads in the Midwest. Why should FSD handle my road differently than a similar road in California?

Any theories?

 

[diplomat33]

Why would this matter? I thought FSD would supposed to be a “general solution” not reliant on HD map data. If that’s correct, why does geography matter? Roads in California are made of the same concrete, lane lines and stop lights as roads in the Midwest. Why should FSD handle my road differently than a similar road in California?

Any theories?

It is not a question of HD maps. Roads and intersections are different based on geography. Vision-only is dependent on the data that is used to train the system. Vision-only can only handle what it is trained to handle. So, if it sees something new that it has not been trained on yet, it won't handle it well. That is not my opinion. Jesse Levinson, an expert on autonomous driving and the CTO of Zoox, says this:

If you build, for example, a vision-only system and then it sees something it has never seen before, there is a very good chance, it totally won't know how to handle it. It might even run into it and not even know it is an obstacle.

That is why FSD Beta handles CA better because Tesla has collected more data from CA. Tesla has less data from the Midwest, so FSD Beta will see things it has not seen before and therefore, will be more likely to fail. That is why Tesla needs more data from areas it has not seen yet.

 

[masterxel]

Slight differences in how the roads appear visually and the background environment can affect neural network perception very easily.

Look how easily a neural network can get the wrong answer even though the image looks identical to a human: How to trick a neural network into thinking a panda is a vulture

They don't bias it towards CA intentionally, but imagine a situation where they can a) remove the CA bias but FSD gets worse everywhere or b) keep the CA bias and do their best to collect more data elsewhere.

 

[nvx1977]

Slight differences in how the roads appear visually and the background environment can affect neural network perception very easily.

Look how easily a neural network can get the wrong answer even though the image looks identical to a human: How to trick a neural network into thinking a panda is a vulture

They don't bias it towards CA intentionally, but imagine a situation where they can a) remove the CA bias but FSD gets worse everywhere or b) keep the CA bias and do their best to collect more data elsewhere.

I probably need some refresher college-lvl math classes before I can truly appreciate that article. Cool to see Karpathy referenced in a non-Tesla setting.

 

[EVNow]

Why would this matter? I thought FSD would supposed to be a “general solution” not reliant on HD map data. If that’s correct, why does geography matter? Roads in California are made of the same concrete, lane lines and stop lights as roads in the Midwest. Why should FSD handle my road differently than a similar road in California?

Any theories?

My theory.

After my fsd experience yesterday - and watching the video I made of my drive - here is my explanation : FSD videos we see on YouTube "feel" different than when we drive. May be if we had VR 360 videos of fsd drives - we would get a better feel. Anyway, definitely actually driving fsd (or earlier just AP) feels quite different (and scarier) than just watching youtube videos - even if the drive itself is similar.

Only way to tell is for you to make a video of your drive - and for us to watch it and compare with other youtube drives.

Note that drives from various places look about the same - Chuck in Jacksonville, Dirty Tesla in MN (?), Frenchie in Chicago, Daily Tesla in WI, and various drivers in CA.

 

[nvx1977]

My theory.

yeah, it's hard to get a sense of the twitchiness of a turn watching a video than actually experiencing it. Today I drove a nice windy country road with tons of tree shadows and the car was lightly applying the brakes throughout. On video, it would look like a perfectly smooth ride.

 

[BigRedDog]

Note that drives from various places look about the same - Chuck in Jacksonville, Dirty Tesla in MN (?), Frenchie in Chicago, Daily Tesla in WI, and various drivers in CA.

Perhaps. I just recall AIDriver or one of the CA testers stating that their drives have gotten boring. My few Wisconsin drives so far have been anything but that.

 

[EVNow]

Perhaps. I just recall AIDriver or one of the CA testers stating that their drives have gotten boring. My few Wisconsin drives so far have been anything but that.

See Green’s tweet. Pretty much the same.

May be it’s better in some CA areas - but more likely the youtubers have got used to all the problems and how to handle them.

Earlier I’d always use AP but my wife was too scared if I used it when she was the passenger. Ask me in a month how I feel about FSD

 

[Dutchy Ron]

I just got the Beta in the “100 Score” wave. I am located in the Midwest. It functions pretty poorly, and far differently than most of the videos I’ve been watching on YouTube.

I think Elon recently tweeted that functionality in locations outside of California would improve as more data from those locations was received.

Why would this matter? I thought FSD would supposed to be a “general solution” not reliant on HD map data. If that’s correct, why does geography matter? Roads in California are made of the same concrete, lane lines and stop lights as roads in the Midwest. Why should FSD handle my road differently than a similar road in California?

That’s the reason why Elons strategy is not the long term solution. Tesla needs in car real time HD maps from TomTom. Better global coverage and receiving more data from different OEM’s.

But the legacy hard- and software platform of Tesla’s is not ready for in-car OTA HD-Maps. Tesla should build a new platform for new cars and models because Elon passed the point of no return with the current platform.

Real time OTA HD Maps will be required for for safety reasons by law soon. Try beta FSD in snow or bad weather. Leve 3 Will be the max for the current Tesla’s

 

[TresLA]

That’s the reason why Elons strategy is not the long term solution. Tesla needs in car real time HD maps from TomTom. Better global coverage and receiving more data from different OEM’s.

It's interesting to see someone so sure of the answer to a future problem not yet solved. I'm guessing from all of your (few) posts that you're paid by or have some financial stake in TomTom. Why not another HD map provider? And by real time do you mean constantly downloading gigs of HD map data? No cellular provider would allow that at this point. Otherwise, current HD mapping goes out of date too quickly. What about dirt roads? Private roads (last mile to many residences/offices)?

The point being, HD maps is not clearly the solution unless you ignore its pitfalls/shortcomings. Tesla and some others aim to have a generalized self-driving system. It doesn't mean everywhere has to be mapped/seen before, but it does mean enough differences need to be seen to be able to better guess the general/new circumstance.

 

[Terminator857]

Elon tested in California as did other Tesla engineers. Every time there was an issue it was fixed. All the data collected was initially from bay area and LA area. Later it was tested outside of California, but has been tuned for California and more specifically the bay area. Most of the positive comments about FSD come from testing in the bay area.

 

[nvx1977]

That’s the reason why Elons strategy is not the long term solution. Tesla needs in car real time HD maps from TomTom. Better global coverage and receiving more data from different OEM’s.

But the legacy hard- and software platform of Tesla’s is not ready for in-car OTA HD-Maps. Tesla should build a new platform for new cars and models because Elon passed the point of no return with the current platform.

Real time OTA HD Maps will be required for for safety reasons by law soon. Try beta FSD in snow or bad weather. Leve 3 Will be the max for the current Tesla’s

I don't see why people think seeing the road is the problem. The hard part about FSD is deciding what to do with what it sees. We humans also struggle to drive through bad weather. The universal solution for that it is to slow way down, so your brain has more time to process what it's seeing and not seeing. What good is an HD map if there's a giant snow bank from a snowplow that just went by?

There's also nothing about Tesla's current approach that prevents it from adopting an HD map if forced by regulators. But enforcing specific implementation details by law is a TERRIBLE idea. We see this over and over with historical bad laws. Example: US headlights (hampered by regulation) vs EU headlights.

 

[Dutchy Ron]

HD Maps know the lines and lanes on the road better than a Tesla which drives in an area where there are not many Tesla’s. Bad weather is I problem for HD Maps because the car can still localize itself with a 10cm accuracy.
FSD including in-car HD Maps are much more safer and I think this will be required by law for L3+.

 

[Jeff Hudson]

HD Maps know the lines and lanes on the road better than a Tesla which drives in an area where there are not many Tesla’s. Bad weather is I problem for HD Maps because the car can still localize itself with a 10cm accuracy.
FSD including in-car HD Maps are much more safer and I think this will be required by law for L3+.

You haven't fully thought out the implementation of FSD. Please adjust your thinking along the lines of what human drivers do in order to perform the act of piloting a vehicle. What decisions do human drivers go through to pilot a vehicle along an unfamiliar road. They do not reference a HD map. They often do reference a map to figure out how to get somewhere but they don't use a map to position their vehicle on the road. Drivers use their understanding of the rules of the road and their driving skills to decide how to proceed at all times and under all circumstances. Most driving problems occur because the task of driving also requires the driver to make assumptions about every thing and proceeding with assigning an order of priority with every thing. Driving is a lot more than merely positioning the vehicle on the road.

 

[Dutchy Ron]

You haven't fully thought out the implementation of FSD. Please adjust your thinking along the lines of what human drivers do in order to perform the act of piloting a vehicle. What decisions do human drivers go through to pilot a vehicle along an unfamiliar road. They do not reference a HD map. They often do reference a map to figure out how to get somewhere but they don't use a map to position their vehicle on the road. Drivers use their understanding of the rules of the road and their driving skills to decide how to proceed at all times and under all circumstances. Most driving problems occur because the task of driving also requires the driver to make assumptions about every thing and proceeding with assigning an order of priority with every thing. Driving is a lot more than merely positioning the vehicle on the road.

I fully agree but the understanding of the road and the exact position of the car are the basis for FSD. Sure, you not only need HD Maps but also need the full stack including AI and camara’s for max safety. Main question is if L3+ without real time OTO HD Maps will work and be safe? You know my opinion ;-). Mercedes will start in Germany next year with L3 and HD Maps.

 

[xEvoLveDx]

My initial drive this morning was borderline terrible and I had to take over several times in a short 4 mile drive. I'm in the Midwest as well (20 miles east of St. Louis in IL). I'm sure this is due to Tesla not having ANY data near me. The roads I just took seem to be marked well but lots of newly constructed lanes and new lights/intersections. I'm wondering if those new intersections/lights have anything to do with the performance of FSD? How do maps play in to this?

 

[Jeff Hudson]

you not only need HD Maps but also need the full stack including AI and camara’s for max safety

When I as a human drive the vehicle I do not need HD maps. I think this is the root of our disagreement.

 

[Jeff Hudson]

How do maps play in to this?

This is the answer Tesla has not publicly answered in detail. I suspect the answer is quite complicated, nuanced and evolving.

 

[TresLA]

At risk of derailing the thread more, here is a good example of why HD Maps doesn’t solve everything (link should start at 9:50 into the video):

Car makes navigation mistake, coned off lanes for construction (i.e. - correct lanes or “map data” has temporarily changed), get stuck, blocks traffic, car moves a couple more times, including running away from the Waymo support tech that is walking to rescue the Waymo, etc.

Even in a relatively controlled and simpler streets system, Waymo’s L4 can have trouble. Hence, Tesla (and others) want to go after a generalized system that can work anywhere.

 

[EVNow]

Tesla needs in car real time HD maps from TomTom.

I've reported you to mods as a shill for TomTom.