Auto-park fail - manually stopped just in time

[R.S]

I can't guarantee that current M3 hardware is adequate for FSD, but it doesn't mean it's never going to happen.

Current EAP is definitely not representative of what can be achieved in the future.

I'm taking a Machine Learning class, cause I want to understand what it's about. I'm like 30% into it and just touched on neural networks, so I don't yet have a full picture. But it does seem very promising. It's basically re-creating brain function by training a "programming representation" of it on a number of scenarios, the same way it happens in life. So, whether it's walking, or driving, or listening to a music, or predicting what's going to happen - you basically accumulate a number of experiences in your life, classify them into these specific categories and then based on the data you accumulated in each category, you learn to recognize/classify/predict.

Driving is just one of many tasks that brain deals with, so you can limit the data you need to feed to the model to this subset of experiences.

It is still a gargantuan amount of data that needs recognized/processed by the model. But you can see how your model is doing in recognizing the data you feed to it and make adjustments, so it becomes better at recognizing your training data.

I think this approach has a good potential in replacing the human at these tasks. But will see what happens.

But google is using neural nets for years now and it's still not working too well (Edit: despite in extensively mapped areas). Despite super precise sensors and all.

So while I do think at some point we will get to real FSD, it's not going to be end of next year. Elon's quotes have always been way too optimistic, he said 2-3 years to 90% FSD, back in 2013. So I see EAP as the most representative indicator we have. There won't be a FSD switch and suddenly the car will perfectly stay in it's lane, recognize every object on the road (stationary or not), plus all the things that are actually way harder, like making actual decisions. Braking for a red light, people crossing the road, left turns at intersections, selecting a lane for an upcoming intersection/off ramp, merging lanes...

 

[JeffK]

But google is using neural nets for years now and it's still not working too well. Despite super precise sensors and all.

Wait what? The Waymo stuff is working great and so far has proven safer than a human. Did you have evidence to the contrary?

 

[R.S]

Wait what? The Waymo stuff is working great and so far has proven safer than a human. Did you have evidence to the contrary?

It works very well in geofenced, well mapped areas, but that's not what Elon plans as a FSD outcome.

 

[JeffK]

It works very well in geofenced, well mapped areas, but that's not what Elon plans as a FSD outcome.

It happens to be currently geofenced. It doesn't require it to be geofenced as far as I know, unlike GM supercruise which requires prior lidar mapping.

Waymo vehicles have their own lidar onboard. We also seen a number of demonstrations where the Waymo vehicle handled strange situations gracefully.

Waymo cars drive like a grandma, but some of the newer MobilEye demos show a more human like driving behavior. If that's where Elon got his inspiration, then I think we're going to be in good shape.

 

[R.S]

It happens to be currently geofenced. It doesn't require it to be geofenced as far as I know, unlike GM supercruise which requires prior lidar mapping.

Waymo vehicles have their own lidar onboard. We also seen a number of demonstrations where the Waymo vehicle handled strange situations gracefully.

Waymo cars drive like a grandma, but some of the newer MobilEye demos show a more human like driving behavior. If that's where Elon got his inspiration, then I think we're going to be in good shape.

The problem is, that they need an extremely good map to work. It has to be 3D mapped to the centimeter, with lot's of tags. It's needed to let the AI focus only on what has changed and therefore greatly reduces the load on the processor.

People often underestimate how much of our human brain is just needed to recognize something, interpret it and filter out the unimportant information. The map tells the AI, what the road is and what's always there, like a building, a street lamp. What it has to look for, a traffic light, for example. And then the car only has to look for differences between it's map and the reality, like a moving car, a pedestrian waiting at the side of a road, maybe a changed speed limit sign. And then it only has to interpret the important/changed data.

This Financial Times article does a great job explaining the need for maps: Driverless cars: mapping the trouble ahead | Financial Times

There is also an interesting quote from Sravan Puttagunta, chief executive of Civil Maps: “That is why autonomous vehicles are geofenced. They physically cannot fit the data in the trunk of the car, so they are restricted to certain areas”

So if you don't want to switch hard drives when you get into a new location, or want to wait for super fast mobile internet everywhere, so you can store it in a cloud, there is no way you could just buy a Waymo car and just use it to drive anywhere, even if everything would already be mapped.

 

[JeffK]

The problem is, that they need an extremely good map to work. It has to be 3D mapped to the centimeter, with lot's of tags. It's needed to let the AI focus only on what has changed and therefore greatly reduces the load on the processor.

People often underestimate how much of our human brain is just needed to recognize something, interpret it and filter out the unimportant information. The map tells the AI, what the road is and what's always there, like a building, a street lamp. What it has to look for, a traffic light, for example. And then the car only has to look for differences between it's map and the reality, like a moving car, a pedestrian waiting at the side of a road, maybe a changed speed limit sign. And then it only has to interpret the important/changed data.

This Financial Times article does a great job explaining the need for maps: Driverless cars: mapping the trouble ahead | Financial Times

There is also an interesting quote from Sravan Puttagunta, chief executive of Civil Maps: “That is why autonomous vehicles are geofenced. They physically cannot fit the data in the trunk of the car, so they are restricted to certain areas”

So if you don't want to switch hard drives when you get into a new location, or want to wait for super fast mobile internet everywhere, so you can store it in a cloud, there is no way you could just buy a Waymo car and just use it to drive anywhere, even if everything would already be mapped.

An interesting example is the MobilEye approach to mapping for autonomous vehicles in which the data itself is measured in kilobytes per mile. It'd be trivial to store entire regions of the country and update them frequently with data collected from the vehicles themselves.

 

[Moderatefan]

Elon's quotes have always been way too optimistic, he said 2-3 years to 90% FSD, back in 2013. So I see EAP as the most representative indicator we have. There won't be a FSD switch and suddenly the car will perfectly stay in it's lane, recognize every object on the road (stationary or not), plus all the things that are actually way harder,

Re: 2013, you do realize that they had to throw away MobilEye AP1 and start over from scratch to re-create the functionality in AP2? That was quite a significant delay/setback.
I honestly don't trust too much that in 2019 we'll have full FSD, that may take couple of years more.
But I do believe they'll be making incremental progress and releasing updates once in a while.

 

[R.S]

Re: 2013, you do realize that they had to throw away MobilEye AP1 and start over from scratch to re-create the functionality in AP2? That was quite a significant delay/setback.
I honestly don't trust too much that in 2019 we'll have full FSD, that may take couple of years more.
But I do believe they'll be making incremental progress and releasing updates once in a while.

Is that why Mobileye has self driving cars now? If they hadn’t thrown out Mobileye, they would have still needed to develop the same skills by themselves. Or leave it all to Mobileye, but they aren’t as far either.

 

[Moderatefan]

Is that why Mobileye has self driving cars now? If they hadn’t thrown out Mobileye, they would have still needed to develop the same skills by themselves. Or leave it all to Mobileye, but they aren’t as far either.

Nobody knows what happened for sure except them. Their explanations don't match Tesla says jealousy, not Autopilot safety concerns, caused breakup with Mobileye

 

[R.S]

Nobody knows what happened for sure except them. Their explanations don't match Tesla says jealousy, not Autopilot safety concerns, caused breakup with Mobileye

Sure, but that doesn't really matter. The point is that neither company has self driving cars now. So it doesn't matter if they broke up, no Tesla would drive itself today.

 

[AmirTech]

So yesterday was my second time trying out a parallel auto park.
The first time worked perfectly, so I knew what to expect.

This was your standard parallel park on a quiet residential street. Lots of space between parked cars, quiet road, parking spot on the right side of the model 3.

As the car was backing up into the space and doing its final "counter clockwise" turn to tuck into the spot, I noticed the front right bumper was getting wayyyyy to close to the truck infront of me, and it was not stopping. I manually used the brakes to cancel auto park at the last minute. The front right corner of the bumper was ~20 - 30cm away from bumping a black truck. I got out to look, and there was no way the auto park would of successfully made it, unless it would of stopped and corrected itself, which it honestly felt like it was not going to happen.

I googled to see if anyone else had a problem, and I found a video of pretty much the exact same thing that happened to me, it was also a truck (only difference was it was on the right side of the car)

Has this happened to anyone else? Kind of disappointed that my auto park is only at a 50% success rate.

Your feedback was really helpful. I am in Toronto, Canada. I wanted to know if Model 3 the basic and standard package comes with Auto park feature? or we have to upgrade to premium upgrade to get have Auto park feature?

Thank you so much

 

[TT97]

Your feedback was really helpful. I am in Toronto, Canada. I wanted to know if Model 3 the basic and standard package comes with Auto park feature? or we have to upgrade to premium upgrade to get have Auto park feature?

Thank you so much

Autopark is part of EAP. It should be available if you order EAP regardless if you have the premium trim or standard trim.

 

[chronopc]

I gave a couple of my co-workers a ride on Friday and they were so impressed by auto park. It may have it's shortcomings but for situations when it works perfectly (i.e. getting into a tight space between 2 cars in a parking garage) it's awesome.

 

[Hammonddave]

AutoPark is definitely a mixed bag. It works 80% of the time for me. On two occasions I had very strange results that could have resulted in collisions had I not been paying close attention.

I live on a one-way residential street that has parallel parking on both sides. One night I was approaching the right side of the street and the “P” indicator came up after I passed a space. I put it in reverse and hit the AutoPark command. But instead of backing up and turning the wheel to the right. It backed up and turned the wheel hard left (as it was trying to park on the left side of the street). Had I not immediately applied the brakes my front right quarter panel would have hit the parked car on the right! It came within a couple of inches.

The other instance is when I approach my garage via the alley behind my house. For some reason the “P” indicator and AutoPark command appear when I stop just past my garage. There is no spot available to park as the garage door is down. I am not brave enough to see what happens if I push the command.

So Let’s just say I am not too confident in in this feature.

 

[jrad6515]

The other instance is when I approach my garage via the alley behind my house. For some reason the “P” indicator and AutoPark command appear when I stop just past my garage. There is no spot available to park as the garage door is down. I am not brave enough to see what happens if I push the command.

So Let’s just say I am not too confident in in this feature.

Mine does the same thing. My driveway off the alley is very short and I need to park sideways in order to wash the car without blocking the alley. Fortunately the autopark function works perfectly for this:



So yes, it can work in this situation.

 

[NickFie]

But my question is, while it is known to not always work, do I want to let myself get used to it and possibly become complacent. I can't fully count on myself here

Aircraft autopilot is also fallible. Pilot/ driver is ultimately responsible for a safe journey.

There’s a category of aircraft disasters caused by pilot over-reliance on the autopilot. I was involved in the aftermath of an incident where cockpit crew relied on autopilot to the point their aircraft fell out of the sky.

Yet NTSB and FAA haven’t banned autopilot use. Within its scope, autopilot manages the vehicle safely and efficiently while reducing crew fatigue.

Through thousands of miles of EAP travel, I’m constantly scanning the road to understand the situation and possible risks. That’s less tiring when EAP manages speed, spacing and lane position.

I also monitor EAP operation. I’ve rarely had to take control. When the occasion arises, I ‘m ready.

EAP can increase safety. It’s still driving while I’m distracted. Once I glanced at the map on my wife’s iPad. When I glanced back up, EAP had gracefully decelerated to accommodate a semi that had moved into our lane.

Also, I don’t have to steer with my knee when I need both hands to open a water bottle.

 

[timk225]

It does say in the manual that you are responsible for it.

The driver has been warned in the manual, and the placement of the ultrasonics is fairly limited, there's not a whole lot that can be done. It's also an assistance feature meant for human supervision, it's not a fully autonomous feature.

Disclaimers like this remind me of modern medicine ads on tv. It fixes one minor little thing, and can cause a huge list of other more serious things. I'll just park my own car like I've been doing for many years now.

 

[StellarRat]

It seems to me that these "super accurate" maps could be derived from collecting the data from cars being driven on the roads by humans over time. If you average speeds, location data, and stop/start info by people driving millions of miles the odds of a bad map are very tiny. I'd certainly rely on that way before I'd trust map info from some municipal database. I still find wrong speeds and impossible routes all the time in the Tesla and Google maps. They certainly would not work for full self driving.

 

[alphawolf9999]

Same thing happened to me Wednesday, we actually hit another car, there are scratches.