Do you have a cite? Because I don’t get it. Identifying other cars is a lot harder to do than identifying the lanes of a freeway.
I assure you, if a Google car on 880 slowed to 2 mph, I’d hear the honks from where I’m sitting.
I can see it - maybe - on a back mountain road with no shoulders to speak of and lots of tree. An unmapped city street - nah. There obviously has to be a high level map for the car to know which way to go, but not that level of detail. For one thing any city street will have a varying distribution of parked cars
Sorry. I’ve been using UNIX system System V, and I spend all my time on Solaris. And I see we agree about Bing.
But if you’ve written any code for people, I’m sure you’ve gotten someone asking if you bothered to check for end of file or null inputs or something. The route at least one Google car drives is the same route I drive to work. It is real. It is awful. The traffic columnist for the Mercury News wrote a column celebrating the fact that he doesn’t have to drive it any more. They are not driving around Mountain View only. They are trying it on a terrible road, and I come from New York and know terrible roads quite well.
They are not ready yet - they know that. But they are way beyond a test track or the desert. The Google car is mixing it up with some of the worst drivers in the US, and I’ve also lived in Boston.
MEBuckner just provided a cite (thankya!). Everything I’ve seen about the car’s capabilities matches that article. I am impressed by them solving the problem that way, I’d been toying with the idea of a crude heads-up display for my car for a few years (I’d just like it to identify a large object, and put a box around it) , but reading about dealing with computer vision made me realize I was in over my head when encountering that problem. Solving that problem is probably going to be job of several careers, a hobby wasn’t going to get far.
Google solving that problem in its limited way provides a method for them to move on while computer vision is improved, but it does limit the usefulness of most of their system to the areas that they can afford (and are allowed) to create and maintain detailed maps of.
Ok, so you understand why I reacted like you’d called me an Uncle Tom :).
I am really impressed by what they’ve accomplished, but the crutch of the map is a pretty significant crutch.
The not being ready anytime soon part of it is clearly correct. But if storing the data is a problem now, in 15 years (which is absolutely the minimum time I’ve heard for a car like this to be ready for sale) it won’t be. Not to mention that the complex analysis done now will by special purpose code can be done for every car and then transmitted - kind of like traffic information is now.
As for AI, while a real AI like Minsky discussed is not a lot closer now than it was then, because they still don’t have a clue as to how to make one, the AI applications I learned about 40 years ago from his associate Pat Winston have pretty much all happened - and happen on our smartphones or PCs. I think autonomous cars are a lot closer to that type of application than to the true AI which I doubt probably more than most people.
If they tried to sell one now, it would be a Newton. But my Android phone understands my speech extremely well and extremely quickly - so not being ready now does not mean not being ready in 15 years.
BTW, any Bay Area driver will tell you that not being able to sense turn signals is not an issue. The few people who use them use them pointing in the wrong direction half the time. You learn to sense what a car is actually going to do, not what it says it will do.
I think it’s a maintenance issue and not a storage/retrieval issue. The storage/retrieval problem appears to have been solved by the current tech, at least in well built up areas. Google can supply a map that it’s scanned, and they can deliver it to their cars.
I did speculate about making all self-driving cars scanners earlier in the thread. You could make it scan constantly. Once it had scanned the route with you driving enough times, it could filter out enough noise and objects to build a map and agree to drive the route for you. Sharing that data presents problems – again, what if you drive your scanning car onto a secure location? On top of that, I think the maps are made with a static* laser, and the cars are using much more low-res lidar. So, having them scan for you may not be as useful an idea as I had originally thought. I don’t think the folks at Google are dumb, and they probably thought of that while mapping their parking lot.
Without having all (or at least a significant portion) of the self driving cars building/maintaining that map for you, I think it’s a maintenance issue and not a storage/retrieval issue. Every time there’s a change in the roadway and the surfaces near it (things like landscaping may present a problem), the map used by the cars will have to be updated. Aside from the astounding amount of roads that even just California or Texas has to map, they change constantly. I have a pretty good mental map of the roads around my home, and I hope I qualify as strong AI, but I have a hard time dealing with the real-time changes sometimes.
It’s not an unsurmountable problem - Google might conquer it with cash, or public opinion might drive a change in automated mapping and how we perform construction. And the ability to drive a car might become similar to security clearances or concealed handgun licenses. Until someone solves real-time visual processing on the level of a human (which for the purposes of driving, possibly could be solved with insanely beaucoup parallel processing power per car, hardware might make it possible), the current tech of self-driving cars is going to be dependent on the map, and how well we can get that map to conform to reality.
You are probably more qualified to speak on AI than I am (I’ve only read in my spare time), and I agree that weak AI has made leaps and bounds in recent years. It seems to me that “knowing” what other drivers are going to do is something that will require either a very clever algorithm I haven’t even heard of, or strong AI. I don’t even really know how I “know” what another driver is going to do, but I often do. I don’t think that kind of intelligence is necessary for a practical all-weather autopiloted car (that description seems to make more sense than self-driving, the more I think about it)
15 years before a commercially available self-driving car with autopilot that can drive most places by itself and park itself with a requirement for an actual driver to be present, seems pretty comfortable. The world could be mapped out well enough by then and the fail safes well developed enough to make it practical.
If we’re talking about a car where I can be morally excepted** from driving long enough to take a nap for the whole trip, I would have this to say: I can’t think of a safe automated system where a human life was at stake where the system didn’t have a human checking in at what is accepted to be a safe interval to make sure nothing is going wrong. In the case of driving, that interval is currently measured in seconds. When we cross the line where it will include a long-distance trip, it will be a sea change, and I don’t have any prediction when it might be or what tech change may allow it to be a good idea.
*The idea of dealing with figure/ground depth perception in a moving vehicle was what made me go “umm, no” with my project. Perhaps with enough low-res scans directly to the left/right (a sweep is a sweep, after all), you could assemble a high-res map. I don’t know if that’s even mathematically possible, before you’d get to engineering the thing.
**And I mean completely, all the way to what is expected of me todat: not getting tickets for being inattentive to a car that had an error and was holding up traffic and crawling at 2mph, to one that was speeding due to the error. We can make safer systems, but abdicating control is a completely separate idea. There are worse crimes than mowing down toddlers and nuns for sport, and I’m meaning actually better than human at this point.
Btw, todat is obviously an ancient contraction meaning “to do that”, and holding up traffic is obviously far worse than mowing down nuns and toddlers for sport. We chase toddlers from horseback for sport in Texas. I dare you to prove me wrong.