Autonomous Car Progress

[daniel]

The fact that HD maps are not 100% reliable does not mean that they are useless. They provide a starting point which, on average, will reduce the time it takes to get to a conclusion and will therefore make the car's decision-making a little bit faster and therefore better.

That's how I read the tweet, anyway.

 

[diplomat33]

I read the tweet, and my comment is in regard to the tweet.
They are stating that - HD maps are useless - we can't rely on them since we have to react to $#!t not in HD maps.
Waste of effort.
The Waymo/Google way!

You are just reading what you want to read. That is not what Waymo said at all.

Can you please give me an example where HD maps (not for geo coded traffic controls and Mid-level maps Tesla uses but High Def) ever trump real-time perception (i.e. vision)??

I'll wait...

You don't get it. It's not a question of using HD maps OR using vision.

It's not about trumping perception. It's about giving the car a priori information. HD maps are like giving a person a picture of the finished puzzle before they start working on the puzzle. You still have to look at the pieces and put the puzzle together (perception) and you can certainly complete a puzzle without looking at a picture first but if you see a picture first, you will complete the puzzle a lot faster.

Yes, I can give you an example. HD maps can tell the car where to stop in the middle of an intersection if they have to wait before completing an unprotected left turn. That is not information you can directly get from vision.

 

[diplomat33]

The fact that HD maps are not 100% reliable does not mean that they are useless. They provide a starting point which, on average, will reduce the time it takes to get to a conclusion and will therefore make the car's decision-making a little bit faster and therefore better.

That's how I read the tweet, anyway.

Yes, exactly. You got it.

HD maps are like looking at a finished picture of a puzzle before starting a puzzle. It gives you a starting point of what it should look like when you are done and therefore will make finishing the puzzle a lot easier and faster.

 

[Tam]

...They are stating that - HD maps are useless...

At a low level of utility, it's useless: For example:

GM Super Cruise compares the current real-life driving data to the pre-mapped HD data and it sees there's a discrepancy, to be safe, it would shut down the system and let human take over because it might be a construction site.

It's useless because Tesla still works with road construction.

At a higher level of utility, it's very useful: For example:

HD data has all the referenced objects such as buildings, trees, poles, bridges... So when there's a discrepancy because the snow has covered up the lane markings, the car can use those pre-existing references that are still accurate between the HD and real-life to triangulate where the lane markings are supposed to be.

It's much more work, but it's much safer and as long as the system is a hard worker (or thinker in this case), it's not useless!

 

[mikes_fsd]

You don't get it. It's not a question of using HD maps OR using vision.

I get it just fine, it seems that your view of computer vision (perception) is very limited in scope.
My view of computer vision - in order to be useful - is that it must recreate the surrounding environment in a very realistic/detailed manner on-the-fly.

But if Rube Goldberg solutions are your thing, then HD maps fit right in.

Have computer vision, stitch it in or diff it with the HD maps have logic to determine which one to obey at which point, then execute your driving plan.

 

[diplomat33]

I get it just fine, it seems that your view of computer vision (perception) is very limited in scope.
My view of computer vision - in order to be useful - is that it must recreate the surrounding environment in a very realistic/detailed manner on-the-fly.

But if Rube Goldberg solutions are your thing, then HD maps fit right in.

Have computer vision, stitch it in or diff it with the HD maps have logic to determine which one to obey at which point, then execute your driving plan.

My view of computer vision is not limited. I fully agree you need vision that can create an accurate 3D perception of everything. I am not suggesting a limited use of vision for only some things and HD maps for other things. I am saying you need both full vision AND HD Maps if you want the highest reliability possible for L5.

 

[mikes_fsd]

I am saying you need both full vision AND HD Maps if you want the highest reliability possible for L5.

You keep treating HD maps as if it can act like a safety net.

1. HD maps cannot be relied on on their own, you agree with that
2. since they cannot be relied on as a backup to a failure in the vision system (which vision should not fail in an FSD vehicle)
   • then you get no benefit of having the extra complexity. (zero benefit)
3. Since vision cannot fail (meaning you need to make it redundant and robust enough to handle component failure) focus on your vision system!

Get your vision in order - since that is an absolute must - get rid of everything else that will slow you down in the long term, even when it gives you a (false) sense of security during development.

 

[Mardak]

I take from this answer that Waymo could self-drive on the sensors alone with no HD map

What happens to something without a "foundation?" Can you proceed to later steps if you're unable to complete the "first step?"

Their response makes it sound like HD maps are critically required and not optional.

 

[diplomat33]

You keep treating HD maps as if it can act like a safety net.

1. HD maps cannot be relied on on their own, you agree with that
2. since they cannot be relied on as a backup to a failure in the vision system (which vision should not fail in an FSD vehicle)
   • then you get no benefit of having the extra complexity. (zero benefit)
3. Since vision cannot fail (meaning you need to make it redundant and robust enough to handle component failure) focus on your vision system!

Get your vision in order - since that is an absolute must - get rid of everything else that will slow you down in the long term, even when it gives you a (false) sense of security during development.

You seem stuck in this mentality of "just solve vision until it never fails and then you won't need anything else". But there is no such thing as vision that never fails.

And yes, there will be cases like construction that will temporarily alter the driving parameters. But the HD maps will still provide reliable foundational information. There is tremendous benefit in having HD maps. They give the car a wealth of reliable information from lane lines, traffic lights, stop signs, legal and illegal paths, landmarks for localization, crosswalks, stop lines etc... that will help you achieve higher reliability of your entire system.

What happens to something without a "foundation?" Can you proceed to later steps if you're unable to complete the "first step?"

Their response makes it sound like HD maps are critically required and not optional.

The fact that they say that their self-driving system can navigate safely if something new or unexpected is encountered, implies to me that their perception could self-drive without HD maps.

Basically, Waymo could self-drive without HD maps but reliability would be lower so they would need more driver supervision. Or Waymo can self-drive with both sensors and HD maps and get much higher reliability where they can remove the safety driver like we've seen in some Waymo rides. Clearly, Waymo prefers the second option. In other words, HD maps are not critical to do "basic" autonomous driving, but they are critical if you want to achieve driverless autonomous driving.

 

[mikes_fsd]

"just solve vision until it never fails and then you won't need anything else"

You HAVE TO... as in MUST... solve vision for FSD.
At one point in the last ~2-4 weeks you agreed with that... but we know how it goes with you agreeing on something...

HD maps do not help if vision is failing. you cannot go through an intersection full of traffic lights on HD maps (or lidar for that matter)
You are adding complexity in the wrong place. You should be thinking and working towards vision being accurate and reliable.

 

[diplomat33]

You HAVE TO... as in MUST... solve vision for FSD.
At one point in the last ~2-4 weeks you agreed with that... but we know how it goes with you agreeing on something...

And I still agree with that. Yes, you need to solve vision problems. But you can solve vision and still find more that needs to be solved. There is always room for improving reliability.

And quit with the insinuation that I flip flop. I don't flip flop. I am not going to fall for your tricks to try to play "gotcha".

HD maps do not help if vision is failing. you cannot go through an intersection full of traffic lights on HD maps (or lidar for that matter) You are adding complexity in the wrong place. You should be thinking and working towards vision being accurate and reliable.

Yes, HD maps can help in some cases where vision fails.

One does negate the other. Yes, you still need to make vision as accurate and reliable as possible. But HD maps help too.

The bottom line is that Waymo and every other company says that HD maps are an important component in making their autonomous driving safer.

So no matter how good your vision is, why would you not want to use something that makes your autonomous driving even safer? You want to deliberately not use a tool that makes FSD safer? That makes no sense.

 

[powertoold]

The bottom line is that Waymo and every other company says that HD maps are an important component in making their autonomous driving safer.

So no matter how good your vision is, why would you not want to use something that makes your autonomous driving even safer? You want to deliberately not use a tool that makes FSD safer? That makes no sense.

The question isn't whether it'll make it safer. Question is whether it's necessary.

 

[diplomat33]

The question isn't whether it'll make it safer. Question is whether it's necessary.

Sure, but safety is part of it. Autonomous driving that is not safe enough is not autonomous driving worth doing.

 

[mikes_fsd]

Sure, but safety is part of it. Autonomous driving that is not safe enough is not autonomous driving worth doing.

Okay then, I am going on record to state the following:

• Because HD maps add to the complexity of any FSD, HD maps make the FSD solution more complicated and thus prone to more (yet undiscovered) issues.
• Because of the above, I believe FSD solutions utilizing HD maps are LESS reliable, and in turn less safe, than FSD solutions that solely rely on solved Computer Vision.

ps: BTW I loved your comment above

 

[daniel]

The question isn't whether it'll make it safer. Question is whether it's necessary.

Actually, the question is whether it allows you to reach the same level of safety at lower cost.

 

[Tam]

What happens to something without a "foundation?" Can you proceed to later steps if you're unable to complete the "first step?"

Their response makes it sound like HD maps are critically required and not optional.

For Waymo, safety is first.

It does homework first with the HD map.

It wants to know where are all the lane markings are. Which ones are faded and which ones are clear.

That's the foundation.

When the real lanes are gone comparing with the HD map, it then must make a decision from known references: median divider on the left, trees on the shoulder on the right... to figure out where to steer the car with disappeared lanes.

Tesla relies on risk-takers. The foundation of safety is not built-into the Autopilot and that's why it's called "beta".

Because it didn't do homework to locate where the lanes are, the car got confused with the gore point and hit the left cement median divider in Mountain View, CA.

Waymo did its homework, the HD map knew where the lanes were, where the gore points were, where the median dividers were. Its system does not get confused to steer into a well-mapped gore point and slammed into a well-mapped divider.

It's just very basic, very elementary, and it's just a foundation!

Tesla took the short cut without the foundation and there have been numerous Autopilot accidents and 3 fatalities.

Yes, you can skip the foundation, skip the homework, skip the hard work, skip the "crutch" but be aware that it may cost the life of a Tesla driver.

 

[powertoold]

Some of the “progress” discussion here is shortsighted to say the least. Tesla has made THE MOST impactful progress thus far when it comes to autonomous cars. All other autonomous car companies are essentially doing simulations and experiments while Tesla is deploying their tech to hundreds of thousands on a monthly basis. The main source of most of our disagreements stems from the fact that we are pigeonholing ourselves into the SAE definitions.

It is silly to think of the different levels of autonomous driving as black and white.

Pointing out some fatalities related to autopilot as rationale to discredit the Tesla fsd approach is silly. This whole safety first thing is silly as well and is akin to the “cancel culture” we’re seeing in many aspects of society. We all know and tesla has shown that autopilot actually reduces accidents overall. That’s what matters when we’re talking about large scale disruptive technologies. We can’t let perfect be the enemy of the good.

 

[Daniel in SD]

It is silly to think of the different levels of autonomous driving as black and white.

Not really. Either the car is responsible for driving or it isn't. It's 100% black or white.

Pointing out some fatalities related to autopilot as rationale to discredit the Tesla fsd approach is silly.

I agree. Whether or not Autopilot/FSD is safer than a human is entirely irrelevant since it is not designed to be autonomous (yet).

 

[Tam]

...All other autonomous car companies are essentially doing simulations and experiments while Tesla is deploying their tech to hundreds of thousands on a monthly basis...

Two different approaches:

Waymo takes responsibility while Tesla passes the bucks to the driver.

Tesla lets owners do the tests such as the smart summon which resulted in minor dings, scratches, and parking lot accidents while Waymo takes control of the test either by its backup driver specialist or the machine itself.

Waymo has been able to have its firefly without steering wheels or pedals drive a blind person around since 2016 so driverless is not something new or unachievable.

...Pointing out some fatalities related to autopilot as rationale to discredit the Tesla fsd approach is silly...

The discussion here is not focused on discrediting Tesla but to see why the "foundation" is needed.

Skipping the "foundation" is fine as long as the user accepts the additional risks.

It's fine to not using the "crutch" but be aware that the user might fall.

It's a matter of consents and the freedom to choose the level of risks.

 

[daniel]

For Waymo, safety is first.

It does homework first with the HD map.

It wants to know where are all the lane markings are. Which ones are faded and which ones are clear.

That's the foundation.

When the real lanes are gone comparing with the HD map, it then must make a decision from known references: median divider on the left, trees on the shoulder on the right... to figure out where to steer the car with disappeared lanes.

Tesla relies on risk-takers. The foundation of safety is not built-into the Autopilot and that's why it's called "beta".

Because it didn't do homework to locate where the lanes are, the car got confused with the gore point and hit the left cement median divider in Mountain View, CA.

Waymo did its homework, the HD map knew where the lanes were, where the gore points were, where the median dividers were. Its system does not get confused to steer into a well-mapped gore point and slammed into a well-mapped divider.

It's just very basic, very elementary, and it's just a foundation!

Tesla took the short cut without the foundation and there have been numerous Autopilot accidents and 3 fatalities.

Yes, you can skip the foundation, skip the homework, skip the hard work, skip the "crutch" but be aware that it may cost the life of a Tesla driver.

I agree with everything you said above, except for one item: Tesla autopilot did not cause those three fatalities: Tesla tells its drivers "This is not a self-driving car. You have to remain alert and take over when needed!"

Refusing to use HD maps will probably result in Tesla being last to market with driverless cars, but may result in Tesla having a monopoly in areas where other companies are unwilling to invest in creating those maps.

And EAP right now is safer for me than a car that does not have it because I use it the way it's supposed to be used: as a driver assist, not as a driver substitute.