-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Why AP 2.0 Won't Be Here Soon, and It Won't Be What You Think It Is
- Thread starter Todd Burch
- Start date
-
- Tags
- Model S
At this point, I think all that is being proposed is the car stop at traffic lights and stop signs, then wait for the driver to tell it to go.
Obviously the current cars don't begin to have the sensor package to clear stop signs and traffic lights, but I think they should be able to recognize and react to them. The current package will also never be able to make 90 turns.
Tesla is getting to the limits of their sensors already for dealing with traffic, and has defenses in place for cars merging into you from some videos I've seen - it was the stationary objects the system wasn't already handling, which is why they are mapping to handle them.
MarkS22
Member
Yeah, this will likely be: geotagging each red light/stop sign, a camera capable of clearly seeing the road markings, inferring the location based on the location of the light/sign, and eventually Car-to-X (Infrastructure) where the lights communicate their current status, such as when they're going to turn yellow and how long before it turns red.
I think this stresses the importance of fleet learning. When you have multiple cars detecting a red light and stopping at a specific point, future cars can look for the light in a specific location based on the GPS data and work more reliably. Personally, I'd expect some form of camera or radar based triangulation to help refine the GPS location data. Not unlike navigating with stars, but using landmarks like a building, store sign, or street lamp to pinpoint the car's location at each red light.
HankLloydRight
No Roads
OK, another scenario I thought about... think about secondary roads with a 45/50 or 55 mph speed limit and stop lights every few miles. In some states I know of (like New Jersey), these intersections have massively long yellow lights. What is the car to do if you're traveling along at 50mph, and the light turns yellow? Most people who travel these roads know that the long yellow lights allows you a lot of time to breeze right through long before a red light. But the car doesn't know that. If the car is close enough to detect a light changing from green to yellow, will it be too close to safely stop, when the driver behind you fully expects you to run the yellow light? At some point the car is going to have to decide that it's actually OK to run the yellow light to prevent being rear-ended or stopping so quickly it's unsafe for the passengers.
And the idea of Car-to-light infrastructure to communicate light status to the car is a pipe dream. The municipalities that have to pay for, install and maintain this equipment aren't going to spring for more complicated hardware so your car can know what your eyes already see. There's nothing in it for them to support that.
RogerHScott
Active Member
I might buy that if the Teslas stopping at the light were always the first in line, but they're not. So there will actually be a blurry region
where they stop, (hopefully) no closer to the limit line than the front of a Tesla, but extending back arbitrarily far, depending on traffic
patterns.
You didn't address my basic doubt about the fleet learning approach: what is the default behavior going to be for lights it has't yet learned
about?
RogerHScott
Active Member
HankLloydRight
No Roads
Sorry -- how close does the car have to be to detect the status of the light? What if the light is already yellow when it is close enough to detect it (meaning, the car doesn't know how long the light has been yellow). If you're traveling pretty fast (say 45mph) and you get close enough to detect a yellow, what does the car do? Slam on the brakes, possibly causing a rear-end collision? Does it decide that there isn't enough room to stop safely on the stop line, so it decides to run the yellow light, which could at any moment turn to a red light, and create another possible accident? (or generate a ticket at a red-light camera?). "I'm sorry judge, but it wasn't me that ran the red light, it was the car driving!"
What I'm saying again, even on straight roads and one traffic light and no turns, there are still plenty of ambiguous situations that will be hard for a computer to decide what to do.
Bet TSLA
Active Member
The argument that "I'm not smart enough to think of a solution to this outlier problem." does not mean that the problem is not easily solvable. And the emphasis on perfection is misplaced -- the system has to be only ~10x better than humans to be viable, a rather low bar. Remind me how many people will die in traffic accidents and how many more will be injured even with a 10x better system. This is not going to be all that difficult (technologically). What's going to be disturbing to people is that its failure modes are going to be different, strange and unfamiliar; so we'll ask how the system could be so stupid rather than being impressed that it is so much safer overall.
Roger, you seem to be implying this is a difficult problem, when it's already been solved by Google and others.
First, I don't see this being accomplished with AP 1.0 hardware. This is a Level 3 task which will require AP 2.0 hardware, which means multiple full color cameras and multiple radars.
Identify:
- Intersection
- Stoplight
- Stoplight status
- Crosswalk or stopping line
- Extents of intersection
When identifying red light, stop at crosswalk or stopping line. If line is not visible or available, stop at intersection extents derived from sidewalks, traffic, etc. If all else fails, warn the user, log the event.
Tell me what I'm missing?
Yeah, also getting a little tired of the implication that all of the autonomy engineers at google, Uber, CMU, Tesla, and everywhere else haven't thought of these "problems" that require about 15 seconds to come up with and post on a forum, and haven't been working on solutions for years. Hell, Uber is about to deploy an in-city system.
No doubt, some of these things are really hard problems. Guess what, smart people have solved a lot of really hard problems.
RogerHScott
Active Member
RogerHScott
Active Member
I'm pretty sure any responsible person at Google, or others, would agree with me that this is, at best, a gross overstatement.
Worked on? Sure. Impressive progress made? Absolutely. Solved? Not even close.
Ok, so here you almost answer the question I posed (note: I was asking a question, not making any assertion): in the (likely, imo)
"all else fails" case, you warn the user and ... keep on going? Stop? Either answer requires, at a minimum, a major departure from
the use model of the current AP, since now the driver has to be ready to "pitch in" -- but not simply "take over" -- at completely
unpredictable times, and potentially frequently.
Don't get me wrong here; I would love stopping at traffic lights to work. But I'm an engineer and I know that there's no correlation
between desirability and feasibility, as much as many people believe/wish there was.
The point of my original question wasn't so much to ask "how would their engineers solve the tough
technical problems" as to ask "what do you folks, as Tesla owners/drivers, want your car to do in these
situations".
Well, you must consider Elon an irresponsible physicist because he's called autonomous driving an essentially solved problem. And, Google's already been driving for millions of miles under city conditions, with it's car stopping at both stop signs and stop lights. So, I'm not sure why you're unnecessarily skeptical of what's being done today. BTW, I'm also an engineer. The situation you outlined is not what I'd personally consider one of the more difficult problems. It's essentially a static image/object recognition problem. I think the more difficult ones will be dynamic, such as how to deal with road crews and or traffic diversions or someone waving a traffic baton.
CarlitoDoc
Member
The obsession with full autonomy baffles me. Having said that, Tesla for now has the edge on semi-autonomous features. I prefer to drive my cars, I do not mind if extra safety features keep me out of trouble.
Bet TSLA
Active Member
And I think the truly difficult problem is dealing with human predators. How do you stop nasties from waving a "traffic baton" at potential victims and having autonomous vehicles deliver them up on a platter? Yes, there are various mitigation techniques, but I imagine people are pretty clever when it comes to outwitting autonomous systems. This will scare people, no matter how much cheaper and generally safer the systems might be.
Full autonomy has the potential to completely transform society and cities and save tens of thousands of lives every year in the US alone, so I can see why people are obsessed with it.
RogerHScott
Active Member
I'd be the first to observe that Elon can be a tad overly optimistic, and maybe even prone to occasional hyperbole? Wow, what an honor!
Canuck
Well-Known Member
It's all about redundancy when it comes to lights and stop signs. Without a second forward facing camera, it's just far too dangerous to allow a Tesla to navigate an intersection based on AP1.0, at least in my view.
Camera-Cruiser
Fully Charged
See, I think you either hit the nail on the head or just missed it.
I think AP, real AP is AI, and I think based on what Musk has said about AI, he knows something we don't, or we may think we know it, but have the order of magnitude off, by well, orders of magnitude.
If the cloud is the brain and all of our cars sensors are providing input, then my car in Southern California benefits from what a car in Michigan experiences. My car will know about driving in ice without me ever driving in it. It won't have to process the millions of miles of ice driving, it will just know what ice is. It will be the same for a millions of things seen by hundreds of thousands, and then millions of Tesla's. It will just know in a neural network way. It will know that piece of tire retread is a non issue and that 87% of other drivers will think the same. But it will take into account that the driver ahead of you has been texting and may overreact to seeing the retread. and will create some distance from that driver.
I may be way off, but what if Musk knows that AI is effectively here. Now the car is driving and it has the aggregate of millions of miles of training to draw from. I can't wait for it to stop because there really is a kangaroo in the road in California, and it knows what one looks like because of Australian drivers.
Last edited:
3Victoria
Active Member
I think we are all talking off the top of our heads. All these problems are being worked on in parallel.
All these problems are being worked on in parallel.Similar threads
