(When) Will EAP or FSD avoid potholes?

[S4WRXTTCS]

It will likely require a mix of things.

There really should be a centralized database of potholes, and of debris in the road.

Where any sort of connected car could report the exact location of the pothole, and debris. Basically like Waze, but where the ADAS took care of reporting things.

When things got fixed the road crew could clear it from the database.

Debris is likely easier for a computer to detect. It just has to be trained on what the camera/radar return looks like for millions of them. I do expect to see some kind of debris detection about a year after HW3 is released.

For potholes I think they can implement major pot hole detection, but for small ones I doubt it. There are so many small pot holes where I live that its like trying to drive through a major land mine field.

 

[lolder]

Tesla is not going the mapping and data base route like everyone else as Musk feels that is too restrictive and subject to outside influences. He wants the car to "see" as humans do because that's who roads are designed for. There will always be potholes that a person or autonomous vehicle can't see or should not try to avoid. The problem is all the low profile wheels on heavy cars and political jurisdictions that won't spend money to repair roads.

Get ready for EV road taxes. They wear out the roads just as fast and we're getting a free ride so far. It wont last.

 

[dww12]

Potholes are just one of many reasons why true FSD is still light years away from reality.

Potholes should be easy enough for machine learning. They are fixed on the road and generally look the same. Also EAP would be faster to know which way to swerve to just barely miss it and not hit a car to the left or right. If a computer can learn to spot skin cancers from pictures at the same rate as a human doctors, pothole detection should be no harder, just need the headroom for it in the computing.

 

[OPRCE]

Perhaps this is the only solution for US roads?

A wrap will improve the aerodynamics, don't worry!

 

[Peteski]

Potholes should be easy enough for machine learning. They are fixed on the road and generally look the same. Also EAP would be faster to know which way to swerve to just barely miss it and not hit a car to the left or right. If a computer can learn to spot skin cancers from pictures at the same rate as a human doctors, pothole detection should be no harder, just need the headroom for it in the computing.

It can barely recognise the difference between a bridge and a truck and the nuances of choosing a line to avoid potholes are way beyond the current system.

I'm sure these things will improve, but I'm not expecting it to advance as quickly as you and I don't think these are trivial problems to overcome

 

[Kant.Ing]

Know what we wish for. Don't want another round of nags or phantom swerve.

 

[Uncle Paul]

Imagine, that sometime in the future, Tesla will be able to real time map items on the road and pot holes from other Tesla cameras picking it up with their video.

A Tesla will drive by a pot hole and the video will see it. They can then transfer that data to other Tesla coming along after them. Same with debris on the road. Kind of like what Tesla is now planning with windshield wipers. Your wipers will come on a few seconds before you get to the rain, as they use data from another Tesla driving just ahead of you on the same road.

 

[Peteski]

Imagine, that sometime in the future, Tesla will be able to real time map items on the road and pot holes from other Tesla cameras picking it up with their video.

A Tesla will drive by a pot hole and the video will see it. They can then transfer that data to other Tesla coming along after them. Same with debris on the road. Kind of like what Tesla is now planning with windshield wipers. Your wipers will come on a few seconds before you get to the rain, as they use data from another Tesla driving just ahead of you on the same road.

What sort of map resolution do you think is realistically possible for locating potholes and debris? The wipers thing is trivial by comparison as it's just a very simple relay action requiring no AI or machine learning. Pothole avoidance however needs both pin-point accuracy (not like the nearest few feet, more like within a couple of inches to be effective) and the intelligence to manoeuvre a safe line around them - which is in itself a very dynamic event.

It's all possible in theory, but not so much in reality. There are many steps to take toward FSD and (IMHO) this kind of fine detail is still way into the future. First we need to master the basics of simply navigating a route safely while following the middle of the lane.

 

[StellarRat]

What IR capabilities? I checked the current narrow view dash cam and it didn't see an IR remote control transmitter lens like almost all smart phone cameras can.

I don't know if Tesla has IR capabilities, however, it can definitely "see" very well in the dark. My guess is that is it does see some IR, but I could be wrong. So, for now, since I don't know for sure, I'll say "low light capabilities". I will tell you that most photoreceptors do pickup IR and that they have to be filtered to toss it out to prevent mis-coloration.

 

[brendon1970]

I don't know if Tesla has IR capabilities, however, it can definitely "see" very well in the dark. My guess is that is it does see some IR, but I could be wrong. So, for now, since I don't know for sure, I'll say "low light capabilities". I will tell you that most photoreceptors do pickup IR and that they have to be filtered to toss it out to prevent mis-coloration.

I don't see why AP couldn't be trained to recognize some types of pothole. If we believe the training mantra, its just a case of showing the NN many positive and negative instances of various obstacles. Once the NN can recognize them (and spatially localize them) then its a problem of where to place the car. Move left/right a little in lane should be doable. Exit lane - well NoA only exists in the US, so no. Brake, maybe but they would want to be very confidant in false negatives otherwise the cars will start stopping for ghosts. Bit of a can or worms.

Lots of interesting negative cases - surface texture, colour, painted markings, childrens chalk drawings, paper bags you name it it will have to be trained.

I think that getting the cars to collect images of potholes could be a long journey. Would need some basic detection capability that might be some small %age reliable, then filter and retrain iteratively until some decent threshold is met. Then think about reactions.

There are some quite uncommon corner cases in Australia where the road is literally a 1000 pothole repairs in close formation (touching each other), one has no choice but to drive over them.

 

[Atari2600]

I would rather it avoid dead animals.

 

[Sammydee]

When will it know what a stop sign is?? There are much more basic things that it's ignorant about besides road hazards

 

[brendon1970]

When will it know what a stop sign is?? There are much more basic things that it's ignorant about besides road hazards

If we believe Elon’s statement that FSD will be feature complete this year, then the advanced HW3 NN probably already does maybe with accuracy limitations.

 

[oddirt]

Just got my first blow out from driving over a pothole in my P3D. Serious bummer! It was dark and I couldn’t see it. Pothole recognition & avoidance sure would’ve helped.

The situation made me think about the ways potholes could be detected by the NN. The system could factor in parallax, the change in potholes’s appearance as the vehicles approach it. The leading edge and the bottom becomes visible as you get closer, something the cameras might be able to detect. This could help assess pothole depth when the system decides a course of action.

 

[argon2018]

This is probably the biggest incentive for me to pay attention to the road while using EAP in stop n go traffic. EAP is completely unaware of pot holes and road debris. Would you really want your FSD to ignore potholes that a small steering correction would avoid? Especially on the 19 and 20 inch rims! Anyone heard anything about this subject? Seems like a somewhat relevant yet undiscussed topic ..

per big boss, FSD will in the future, prob after they release street light and stop signs

 

[SammichLover]

Perhaps this is the only solution for US roads?

View attachment 379900

A wrap will improve the aerodynamics, don't worry!

*cough*carbon fibre spoiler*cough*

 

[afadeev]

This is probably the biggest incentive for me to pay attention to the road while using EAP in stop n go traffic. EAP is completely unaware of pot holes and road debris.

Yes, obstacle awareness should definitely be more aware of road debris. I was on EAP and luckily took over just in time to not run over a ladder that was in my lane. You really have to be aware while using EAP.

I would just settle for EAP to stop trying to impale me on the lane divider at the Y-intersection - it always gets lost, and steers right down the middle.
Every time - same intersection, same outcome.

Know what we wish for. Don't want another round of nags or phantom swerve.

+1 on that.

We are having a hack of a time getting our Tesla's to track straight, without random braking and self-impalement on infrastructure objects.

You sure you have enough diapers to let it experiment with swerving ?!?!!?

 

[Watts_Up]

If FSD and RoboTaxi are coming starting next year (Elon Musk promises fleet of 1 million Tesla ‘robotaxis’ in 2020)

then detecting pot holes and speed bumps will be crucial!

Speed Bumps Help Kill 25,000 People a Year in England, Health Group Claims​

 

[Leafdriver333]

Wheel and tire companies love potholes.

 

[Rottenapplr]

Potholes are just one of many reasons why true FSD is still light years away from reality.

And objects that are on the road it can’t see! That kinda freaks me out. I don’t want my car to just barrel through an object the side of a dog for example or a fallen bumper or trash can. You see all kinds of stuff on the road.