Elon: "Feature complete for full self driving this year"

[diplomat33]

I'm waiting for the 1 million hour uncut demo video. Then I'll be convinced that L4 is close!

And you are going to sit there and watch the entire thing. LOL.

 

[emmz0r]

I don't think HD maps are useless. HD maps can help your car "see" something that is blocked or help your car when your sensors fail. As a result, the industry actually considers them to be necessary for autonomous driving.

Here's what the "Safety First for Automated Driving" paper has to say about HD maps. This might help you better understand their role:

"An in-vehicle map has never played a safety-related role as it could do in automated driving. For a relatively long period of time, the capabilities of onboard sensors alone will be insufficient to meet the high reliability, availability and safety requirements of the automated vehicle system in certain situations. A HD map is therefore necessary as a reliable off-board sensor containing carefully processed a-priori information to “detect” features that are not easily detectable by on-board sensors or to provide a redundant source of information for on-board sensors, including location-based ODD determination, environment modeling in adverse conditions and precise semantic understandings in complex driving situations. In situations where on-board sensors cannot reliably detect features, the HD map can be utilized as a more reliable redundant source of information."

Yeah it could complement the sensors but then it has to be updated 24/7. Sounds like a lot of effort for something little extra.

I'm waiting for the 1 million hour uncut demo video. Then I'll be convinced that L4 is close!


You rely on HD maps to improve the performance of the system. For example Tesla maps overpasses to reduce phantom braking. Sure they could just slam on the brakes every time they approach an overpass but that would make Autopilot useless and unsafe. They could just never brake on stationary radar return (like old school adaptive cruise control) but that would also be less safe.

It doesn't phantom brake on overpasses (anymore). I have tons of them, it phantom brakes because the speed limit in the ramp near the overpass is read instead of on the highway. That's what happens to me. When they do away with the stupid GPS speed limits it will be much better. Also it randomly brakes if there is no speed data in a "tile". Very stupid. I wish I could disable it and the speed limit was display only.

 

[diplomat33]

Yeah it could complement the sensors but then it has to be updated 24/7. Sounds like a lot of effort for something little extra.

No, I don't think HD maps need to be updated that frequently. HD maps would only need to be updated for things like a new lane is added or a new traffic light is added or such. That would not be 24/7.

 

[mongo]

No, I don't think HD maps need to be updated that frequently. HD maps would only need to be updated for things like a new lane is added or a new traffic light is added or such. That would not be 24/7.

If the map needs to be updated, what happens in the time between the road change and the update?

 

[Daniel in SD]

If the map needs to be updated, what happens in the time between the road change and the update?

The car uses the other sensors and is probably less safe until the map is updated (which should occur automatically as more cars encounter the change). Of course if it senses a discrepancy between the rest of the perception system and the maps it can go into a more cautious state (perhaps lower its speed or be more conservative on its path prediction of other cars) so maybe there would be no degradation of safety but it would be more likely to get stuck and slow down traffic.

 

[mongo]

The car uses the other sensors and is probably less safe until the map is updated (which should occur automatically as more cars encounter the change). Of course if it senses a discrepancy between the rest of the perception system and the maps it can go into a more cautious state (perhaps lower its speed or be more conservative on its path prediction of other cars) so maybe there would be no degradation of safety but it would be more likely to get stuck and slow down traffic.

Thus the question: did the map really need updated?
Either the car is unsafe without the update and cannot be driven on an out of date road section ( which gets back to needing instantaneous updates). Or the car is sufficiently safe without the map and thus updating it is not a requirement and the map itself is not needed (beyond navigation).

I agree that maps can improve performance over not having them, but if they are put in the 'need' category, then the system is too brittle to operate well.

 

[Daniel in SD]

Thus the question: did the map really need updated?
Either the car is unsafe without the update and cannot be driven on an out of date road section ( which gets back to needing instantaneous updates). Or the car is sufficiently safe without the map and thus updating it is not a requirement and the map itself is not needed (beyond navigation).

I agree that maps can improve performance over not having them, but if they are put in the 'need' category, then the system is too brittle to operate well.

Safety is not a binary thing. It's very likely that the first vehicle to be "sufficiently" safe (statistically safer than a human is the consensus) will use HD maps. It may be possible at some point after that to make a vehicle "sufficiently" safe without HD maps but why would anyone do that if keeping HD maps makes it even safer? Unless the cost of maintaining the maps is higher than the expected cost of additional accidents it seems like they would just keep using maps.

 

[mongo]

Safety is not a binary thing. It's very likely that the first vehicle to be "sufficiently" safe (statistically safer than a human is the consensus) will use HD maps. It may be possible at some point after that to make a vehicle "sufficiently" safe without HD maps but why would anyone do that if keeping HD maps makes it even safer? Unless the cost of maintaining the maps is higher than the expected cost of additional accidents it seems like they would just keep using maps.

My feel is that maps let the car drive faster safer by knowing when to drive slower. Current drivers don't take in account line of sight, hill crests, or blind curves very well. Will we expect FSD to drive as quickly/ recklessly as we do, assuming there are no obstacles? Or will we expect it to drive safely where it would never travel faster than the visibility allows stopping in?

 

[AlanSubie4Life]

Will we expect FSD to drive as quickly/ recklessly as we do,

I hope so. Not going to be a lot of interest from Californians if the FSD vehicles won't drive well in excess of the speed limit. All very theoretical and years in the future, but should be ok, with very fast reaction times. Maps would be helpful of course - humans use their internal maps for the same reasons you outline.

 

[Az_Rael]

I hope so. Not going to be a lot of interest from Californians if the FSD vehicles won't drive well in excess of the speed limit. All very theoretical and years in the future, but should be ok, with very fast reaction times.

If the car company is liable during FSD operation, I can't imagine one authorizing speeding while on FSD use. I suspect real FSD cars are going to drive following every rule, but you won't care since you will be asleep or reading a book or whatever.

 

[mongo]

Intrinsically safe driving would not be popular, so how does one define an acceptable level of unsafeness?

 

[diplomat33]

Intrinsically safe driving would not be popular, so how does one define an acceptable level of unsafeness?

Who cares if it is not popular? It would be dumb to make autonomous cars less safe on purpose just to "fit in" with how unsafe human drivers like to drive.

 

[AlanSubie4Life]

It's not clear to me that driving the speed limit is safer than exceeding the speed limit, if most cars are traveling at 10-15mph over the limit. Obviously in the distant future all cars would be FSD so that solves that issue.

In any case speed isn't necessarily inherently unsafe - it depends on the conditions of course (hence maps to help with identifying more dangerous areas). And obviously with instant reactions you can accommodate a lot of behaviors that a human could not. Not saying that those behaviors would be rolled out en masse, but certainly they aren't the primary obstacle to making FSD vehicles "much safer than a human driver." That stuff is just details and fine tuning around the edges of the incredibly difficult problem.

Anecdotally, the people I see driving the most slowly seem to be part of the least safe driving cohort.

 

[emmz0r]

I think it's actually more important with negotation and social interaction in traffic, how will that work out? Mobileye seems to be on this, while it seems entirely lacking at Tesla.

 

[Daniel in SD]

My feel is that maps let the car drive faster safer by knowing when to drive slower. Current drivers don't take in account line of sight, hill crests, or blind curves very well. Will we expect FSD to drive as quickly/ recklessly as we do, assuming there are no obstacles? Or will we expect it to drive safely where it would never travel faster than the visibility allows stopping in?

I don't think a self driving car will ever drive faster than its vision. However knowing what's coming could certainly make it drive more smoothly and avoid some emergency braking!
In terms of safety just think about stoplights. If you know where all the lights are you can much more accurately detect red lights. Let's say without maps you can accurately detect red lights 99.9% of the time. However if you know where the light is you don't have to rely only on red light detection, you can now make sure that the light is green before crossing the intersection. The chances of falsely detecting a green light are probably much smaller than the chances of not detecting a red light or a light that is broken.

 

[conman]

It never ceases to amaze me how you can pick any page out of this thread, and a dozen other threads over the course of the last 3 years, and not be able to tell them apart. It's like groundhog thread.

 

[Az_Rael]

It never ceases to amaze me how you can pick any page out of this thread, and a dozen other threads over the course of the last 3 years, and not be able to tell them apart. It's like groundhog thread.

Eh, it is a lot like the threads while we were all waiting for the Model 3 to be released, everyone rehashing the same subjects for 2 years. Good times. Once Tesla releases something it will get more interesting I am sure.

 

[AlanSubie4Life]

It never ceases to amaze me how you can pick any page out of this thread, and a dozen other threads over the course of the last 3 years, and not be able to tell them apart. It's like groundhog thread.

Nature abhors a vacuum.

 

[Daniel in SD]

It never ceases to amaze me how you can pick any page out of this thread, and a dozen other threads over the course of the last 3 years, and not be able to tell them apart. It's like groundhog thread.

Yeah, it's really kind of silly.
Has anyone attempted to evaluate the "FSD preview" in terms of detection distance, false positive and false negative rates? That would be something new...

 

[AlanSubie4Life]

Has anyone attempted to evaluate the "FSD preview" in terms of detection distance, false positive and false negative rates? That would be something new...

The results would be immediately invalidated by claims that the preview only shows a fraction of what the super secret code is really seeing.