Yeah, a human driver could fuck up too. If that really was auto pilot, it certainly did fuck up. Look at the road. The conditions where not that bad. Not at all. Now if the car had just regular street tires on it, they might be a little rough.
The car did recover from it’s fuck up. Would be better if it didn’t screw up in the first place.
I am completely unimpressed at how ‘well’ the car did, since it caused the problem in the first place.
That’s not how it works. The car looks for landmarks. These landmarks are very small - they are generated from earlier stages of the vision pipeline. Like a landmark could be “letter e next to letter r” (a piece of a sign visible from the road". If NOWHERE ELSE in the rough geographic area determined by GPS a car has seen this intersection point, that would be a valid landmark to store. And you can imagine that the landmark will be just a few bites and as well there would be a hash of the landmark descriptor making the lookup time O(1). At any given location you want to see dozens of landmarks at a minimum because this makes the car resistant to the weather and human activity changing how the surroundings look.
You don’t want to have to look them up via wireless either. Not only does that cost money but it might fail. Instead you want the car to store the latest area map, updated daily, including landmark updates from other cars in the same network, and to download any segments needed for a trip before leaving the charging/storage lot. (It would obviously download by wifi)
I would have to agree. I am in GIS. GPS will tie you to the closest real mapped road network on the ground. That’s what it displays to you in your car. A ‘you must be here’. Now those are getting more accurate every day, but still…
And differential GPS and true GPS survey stations take a fairly long time on static base stations for true accuracy.
Why is it obvious that the car would use WiFi? I would think a cellular standard, like 5G, would be more likely.
Because that’s how Teslas already do it. And it’s much cheaper. 5G isn’t needed. Nor is it needed for V2V, it’s actually a non-starter, the latency is too high and the bandwidth demands of V2V is also too great. (there are V2V prototypes using 4G/5G, my company is working on them, but ultimately there’s a short range RF band reserved just for this that makes a lot more sense to use)
OK, I’m googling and find that you’re correct that Tesla updates the maps over WiFi. But a lot of the links are from people who don’t have WiFi access where their car is parked. So I don’t think that’s the best long-term answer.
Yep: new technology wins pretty much every time.
…except when it doesn’t.
That video is just stability control doing what it’s supposed to do. The same thing happened to me last year, except I was in control of my car, up until the point I lost control of my car. I was going too fast for the conditions (unexpected ice on an otherwise dry road), and before you know it, I was going sideways wondering how long I’d have to wait for AAA to show up. Just a second or two later, the car righted itself, and I continued on my merry way (a lot slower). This was a run-of-the-mill, late model Mustang.
I thought it was magic at the time. I didn’t know what “stability control” was, other than something mentioned as a feature on the window sticker. It so happens that on that day I was going to the plant where it was built, and I told the guys who built it how it’d saved my life, like magic, and after laughing a bit they schooled me on the latest car technologies.
Sure. But if we can believe the video, it was the auto pilot that put the car into that problem in the first place. As another poster said, it looks like the left fog line was obscured causing the car to turn rather sharply to the right and lose control.
My Wife and I also have VDC on our cars. It does pretty good, but needs to be turned off for deep snow.
In 50 years you may not have to worry about too much snow…
So what do you think makes this challenge difficult for AI, other than simple market size?
Have you plowed much snow? This is going to be very difficult to explain to someone that doesn’t have a LOT of experience doing it, but I’ll give it a shot.
First, 30 feet of snow a season is pretty average. We have had 12 feet so far with March being our snowiest month still on the way. We will have snow in the yard until June.
I’ve been plowing my driveway (and sometimes the county road) for 26 years. The driveway stays the same shape, NO storm is ever the same though. I have a general plan. Move cars out of the way (if they are in the driveway), do the top part first and make a big swipe to push snow towards the edge. Now the edge gets pilled up, and I do my best to push the snow as far as possible but these eventually turn into snow storage. Hard packed snow that the truck will not move and I have to get on my 4x4 Kubota to move, breaking it up and pushing it a bit more. Sometimes carrying the snow to a different spot. If the tractor gets stuck, I pull it out with the truck (I have a winch on the back, an odd set up, but you can’t put a winch on the front of a plow truck).
After the first push of snow on the top part, you have to keep working it depending on the snow conditions, and just how much snow is packed on the surface. Is it a muddy spring? Is it -10 below? It all maters. And is never the same.
After the top parking area is done, I work on the drive itself. This is usually not to bad, but also takes a plan. I have to move as much snow as I can without sticking the truck. I get one shot going down the drive and have to move as much snow as I can in forward and back motions to make sure the truck will make it back up to make another run at it. After 26 years of doing this, I’m still relieved when it’s done and I don’t have problems.
The truck is an 06 Dodge short bed regular cab. The shorter wheelbase gives me a tighter turning radius. It has tire chains on all 4 wheels (aggressive ones that warn not to be used on a paved surface). It has a 10,000lb rated winch on the back that I have a pully block for to change the direction of pull. And yes, it has been stuck. Ruins your day. Toying with the idea of putting a winch on my 4Runner, so I have more options.
If there was enough demand for it, sure, something could be built. I could see a smaller skid steer vehicle taking care of a daily type plowing when my wife and I are away. Plow when ever a snow fall adds up to say, 3 inches. It’s still gonna need help with opening up snow storage areas though.
Programming a vehicle to plow a parking lot would be childs play compared to this.
What I’m trying to imply is that the actual snow plowing is like a 1% challenge. That is, 99% of the work - getting to rock solid reliable autonomous cars, with the behind the scenes hardware, hardware device drivers, virtual machine sandbox layer, the machine learning architectures, meeting hard realtime requirements, the sensors, maintenance of the vehicles -
Basically assuming you already had mass market autonomous cars, they just can’t plow snow, but you have a stable platform both software and hardware and it’s architected sufficiently loosely that it can be reused (not a guarantee), snow plowing dynamics are actually fairly easy by comparison.
Even if they are highly unpredictable (stochastic) or complex, compared to already demonstrated AI algorithms solving many such challenges, it’s not a difficult problem. Basically the problem space is a discrete grid covering the street and your driveway. The machine knows the approximate snow depth from prior mapping data. The goal of the planner is to device a series of manipulations where at the end of the series of manipulations, less snow is on the discrete grid you care about. This is pretty straightforward and I am pretty confident a number of modern machine learning algorithms would fairly easily converge on a stable solution.
My real world job is GIS. Spatial Analysis. We compare any feature to any other feature. It’s really the ultimate data base because EVERYTHING (except perhaps thought) has one thing that can be related to another. That thing is its location.
I was in GIS before that acronym came into use. Used to be called AM/FM. Very confusing. That stood for Automated Mapping and Facilities Management. My staff and I manage road networks for routing for EMS. Among other things.
Now high precision localized GPS with special equipment (like I said like a skid steer) may someday be able to plow snow in difficult situations such as where I live. With help.
I asked you earlier, how much snow plowing experience do you have?
Doesn’t snow down here. I assume the behavior of snow is similar to that of raking leaves. And I’m not sure what GPS has to do with anything, GPS isn’t the primary method that autonomous cars use to determine position…
A Waymo user speaks:
Maybe kind of almost similar to raking leaves. Leaves however don’t generally pile of to the point of being immovable. Nor do they get you stuck.
So. No GPS? Dead reckoning? Visual cues?
Dead reckoning won’t work (IMHO) for a vehicle that moves back/forth and turns going over bumps and rapid starts and stops 100 times in one hour (I just plowed my drive if you can’t tell) If you are off by a foot, you can be SOL (been there).
Visual cues? Maybe. provided you can keep them high enough out of the snow and out of the way. You’ll need to have three at any time to triangulate off of. And the sighting system doesn’t constantly get obscured by snow and dirt.
A VERY localized high accuracy gps, with multiple stations would be the best bet.
I’m not trying to be snarky. Remember, I’m not talking about a standard public roadway. I’m talking about a personal property that gets up to 30 feet of snow a year.
There are robots that use a random approach similar to a Roomba for keeping your lawn mowed. Perhaps you might use a similar approach to automate keeping your driveway free of snow? I think you would start it up while there’s only a couple of inches of snow on the ground, rather than waiting for a foot of wet, heavy snow.
30 feet on average? Aspen averages about 15. I thnk you are living in such extreme conditions that it’s reasonable to guess you’re never seeing autonomous cars.
Right now we are basically retrofitting. Maybe it’s many decades away but I gotta think that in the future, the infrastructure on roads and highways will include beacons of some sort that will augment self driving cars internal abilities to understand where it is.