Another reason why Tesla is miles off full self driving

[Nuclear Fusion]

Part of a chopped up fallen tree was jutting out from the side of the road (into the paved road area) & I got distracted for a couple of seconds while on autopilot & then saw it at the last millisecond, swerved but manage to collect it such that it’s broken the front passenger wheel arch trim. The swerve mitigated the damage, so glad that I didn’t hit it straight on.
Yes, it’s ultimately my fault for losing concentration, but it’s going to be quite a while before the car can recognise such things amongst all the background. It’s why Level 4 might be 3 years away, but Level 5 is minimum a decade away

 

[Gwgan]

Why is this AP/FSD thread in the Model X forum?
The AI driven, video based FSD software that will be running on the car for Level 5 is not the same software that could not interpret and avoid that the tree limb so that’s not an appropriate extrapolation. I’m not saying that level 5 is closer than 10 years, only that this event is not a good justification for that timeline.

 

[_jmk]

This is about the fact that they have to teach the machine vision absolutely everything that can ever be on your way before its ready. This is not realistic. Currently it recognizes probably dozen things at best.

 

[mswlogo]

Neither AutoPilot or Humans will be perfect. But Elon is betting AutoPilot can statistically be better than average humans soon. It has some disadvantages over a human but it has many advantages too.

It never gets distracted. (OP post is a perfect example)
It has great response time.
It never gets tired.
It has quite a few sensors.
It can multitask better.
It is better connected to the vehicle.
It knows the limits of the vehicle better.
With all the above It can calculate the perfect response and adjust it way faster than humans.

It’s just not smart enough yet. But with those advantages it may not have to be as smart as a human to statistically drive better.

Think of the crash prevention features of cars today including Tesla. It clearly shows in these limited situations it can do better. One feature in crash prevention I had on my Jeep was really cool. If it saw I was approaching an obstacle to fast it would alarm me and it would preset the brakes for quicker stopping. If the driver hit the brakes it was confirmation the vehicle was in trouble. So if the driver was not hitting the brakes hard enough the crash prevention system would take over. And it could calculate exactly how much braking was needed and how much the vehicle could withstand.

 

[_jmk]

The human brain has the fundamental advantage that by the age of 16 it has been training identifying and evaluating distances to different objects for 16 years, 16 hours per day. The engine behind the operation is a product of a long evolution and even a baby does a far better job at this than any system out there now. The brain is capable of building a near-perfect 3D model of the surroundings that updates with millisecond latency and with very high reliability. Anything but and you'd be dead by now.

Elon seriously underestimated that effort by the factor of 10-100. Lidar is the only short term easy way out. At this time he is just mostly trying to buy more time.

 

[forkee]

This doesn't even begin to scratch the surface of the bigger philosophical issue known as the "trolley problem"

even when the technology has advanced to where we need it to be, political decision making and law drafting will hinder any true societal integration due to the ethical dilemma of not just driving, but making safe choices along a route during dangerous situations.

 

[CyberGus]

I, for one, welcome our new robotic overlords

 

[DoctorVenkman]

The human brain has the fundamental advantage that by the age of 16 it has been training identifying and evaluating distances to different objects for 16 years, 16 hours per day. The engine behind the operation is a product of a long evolution and even a baby does a far better job at this than any system out there now. The brain is capable of building a near-perfect 3D model of the surroundings that updates with millisecond latency and with very high reliability. Anything but and you'd be dead by now.

Elon seriously underestimated that effort by the factor of 10-100. Lidar is the only short term easy way out. At this time he is just mostly trying to buy more time.

Yeah, it's an admirable goal and I hope we do end up seeing vision-only level 5 self driving. But if anyone thinks we will see real FSD (not the gimmicks Tesla calls FSD) within the next five years they are seriously drinking the kool aid. More realistically I think it's at least 10 years out.

It's just really, really hard to create an accurate and precise 3d mapping using only vision. Without that 3d mapping there's no way to safely navigate the "march of 9's" that Elon likes to reference - i.e. less and less likely edge cases that need to be safely navigated.

 

[AlexF1]

Yup it's going to some time to go IMO. I never really had a complete trip I totally didn't have to intervene even on the highway, not to mention all the city roads.

 

[ChrML]

This is about the fact that they have to teach the machine vision absolutely everything that can ever be on your way before its ready. This is not realistic. Currently it recognizes probably dozen things at best.

You have the wrong approach to the software problem, doing it like that is a really bad idea.

What you want is let the software figure out whether the area is flat/drivable or not, and map that area in memory. You don't need to be able to recognize every single object possible, but you need to tell it's not flat. There exists vision to 3D software already capable of half that task. Hence if an UFO lands on the road some day, your self-driving car will stop or go around it because there's an object in the planned path.

 

[Nuclear Fusion]

Elon seriously underestimated that effort by the factor of 10-100. Lidar is the only short term easy way out. At this time he is just mostly trying to buy more time.

The guy is seriously dumb. He states at the last corporate earnings that “on my alpha build I can nearly get to work without a single intervention, even allowing for roadworks”. Whoop-di-fu**en-do.
Elon, you need to grasp where the number 9 is ranked when you talk about the march of 9(99999999999999999)s

 

[KentuckyX]

Yeah, it's an admirable goal and I hope we do end up seeing vision-only level 5 self driving. But if anyone thinks we will see real FSD (not the gimmicks Tesla calls FSD) within the next five years they are seriously drinking the kool aid. More realistically I think it's at least 10 years out.

It's just really, really hard to create an accurate and precise 3d mapping using only vision. Without that 3d mapping there's no way to safely navigate the "march of 9's" that Elon likes to reference - i.e. less and less likely edge cases that need to be safely navigated.

Agreed. When I got my Model 3 in early 2018, I thought it would be a goal that would be achieved within a couple of years. But here we are in the latter half of 2020 and EAP still can't read Speed Limit signs, avoid pot holes, figure out what to do when the right white line disappears at an exit, self park (worth a crap), stop panic braking or auto lane change balking on the freeway with no vehicles anywhere in the vicinity, or (fill in your own shortcomings here).

 

[_jmk]

You have the wrong approach to the software problem, doing it like that is a really bad idea.

What you want is let the software figure out whether the area is flat/drivable or not, and map that area in memory. You don't need to be able to recognize every single object possible, but you need to tell it's not flat. There exists vision to 3D software already capable of half that task. Hence if an UFO lands on the road some day, your self-driving car will stop or go around it because there's an object in the planned path.

Probably, but it it would still need to reliably detect the object from the surrounding noise on every imaginable road on the planet. Not easy. Sure you can demo this in sunny California, but..

 

[hgmichna]

The guy is seriously dumb. He states at the last corporate earnings that “on my alpha build I can nearly get to work without a single intervention, even allowing for roadworks”. Whoop-di-fu**en-do.
Elon, you need to grasp where the number 9 is ranked when you talk about the march of 9(99999999999999999)s

Looking dumb, but laughing all the way to the bank.

I think he knows. His optimism has a purpose. If he were in fact dumb, he would not be leading a company that builds and operates spaceships that actually work and auto-dock at the International Space Station.

 

[Nuclear Fusion]

Looking dumb, but laughing all the way to the bank.

At his peak of cocaine trade Pablo Escobar had about $60B in wealth, so a pretty smart guy too...

 

[DoctorVenkman]

Elon is pretty smart, but he's no genius from a technical standpoint. His real talents are taking (calculated) risks, motivating people smarter than him to work hard, and getting people buy into his vision. I wouldn't take anything he says about self driving with more than a grain of salt - he has a habit of taking something one of his engineers said and hyping it beyond reality.

 

[Nuclear Fusion]

(Elon) has a habit of taking something one of his engineers said and hyping it beyond reality.

Bingo. The sad part is what that engineer cops behind the scenes from Elon, borne out of ‘being given enough rope’
I’ve read/heard enough interviews to see what a psychologically fragile individual Musk is

 

[hgmichna]

Bingo. The sad part is what that engineer cops behind the scenes from Elon, borne out of ‘being given enough rope’
I’ve read/heard enough interviews to see what a psychologically fragile individual Musk is

More than you think is calculated.

 

[ItsNotAboutTheMoney]

This doesn't even begin to scratch the surface of the bigger philosophical issue known as the "trolley problem"

<snipped because sometimes I hate philosophers.>

even when the technology has advanced to where we need it to be, political decision making and law drafting will hinder any true societal integration due to the ethical dilemma of not just driving, but making safe choices along a route during dangerous situations.

The trolley problem is a ridiculously low probability event that doesn't have to be solved.
You just use the following basic rules:
1) Obey traffic laws
2) Try not to hit people, or other things that will damage the car
3) If you must hit things, do it as slowly as possibly.
4) Record everything

If a human driver, obeying the law, is faced with an inevitable collision that is not their fault, chooses to hit the brakes and not change direction, they are not held liable. The same would be true for an autonomous vehicle.

So, just make the vehicles measurably better than typical assisted human drivers, and stick to simple rules in collision and there's no problem.

The point of the trolley problem is not to solve it, it's to point out that some situations have no right answer, and that people's behavior will depend on how it is framed, and particularly that people are likely to choose not to choose.

The bigger problem will be people messing with AVs, knowing they can force a stop.

 

[_jmk]

Elon is pretty smart, but he's no genius from a technical standpoint. His real talents are taking (calculated) risks, motivating people smarter than him to work hard, and getting people buy into his vision. I wouldn't take anything he says about self driving with more than a grain of salt - he has a habit of taking something one of his engineers said and hyping it beyond reality.

That's absolutely right. He has the correct high level strategic view as he listens to the right people and the execution is top notch. He is nearly the perfect leader when it comes to delivering results. The only obvious mistakes he does are when he tries to improvise too much, like with the Xs design, but he has clearly learned from these mistakes.