Years long battle with a patent troll over core Facetime technologies, don't think it's Apple's fault.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Neural Networks
- Thread starter lunitiks
- Start date
The battle was over NAT traversal technology, and more fundamentally FaceTime is a proprietary stack with Apple controlled presence servers and Apple-managed iMessage encryption keys. There was no way this was architected to be an open standard.
Didn't need Apple servers if you can traverse NAT's. Also you need something for key management, it's surprising keys are managed by Apple's own infrastructure???? This means nothing.
Right because all open standards start with one company that owns the root signing certificate hardcoded into all the devices that support it. And BTW there are many open standards that cannot be implemented due to patent encumberment, that doesn’t stop you from specifying an open standard. All of the MPEG codecs for example.
At any rate it doesn’t change that it’s somwthing announced that the company clearly did not have a finalized plan or how to productize it. Not all announcements are made when it’s a slam dunk path to shipping it.
The big boys (Waymo et al) look for both position and color, but they have HD maps and centimeter-scale localization within those maps, so they know exactly where in 3D space each bulb is, which lane(s) that bulb controls, and what the configuration of that bulb assembly is (horizonal, vertical, or any number of weird others). Tesla does not have this benefit. Tesla approaches each intersection de novo.
Even for humans this is sometimes an issue -- I know that when I'm approaching an unfamiliar intersection I am often confused by the signaling. But at least I have the ability to learn that intersection from a single example; unless Tesla steps up their game on labeling in a very serious way, they are not going to be able to learn from single examples. And for them to turn their cute little open-source road maps into HD maps is a massive effort, and even then they have no reliable centimeter-scale localization. (GPS doesn't cut it -- in theory camera localization is possible but nobody has shown it at any scale yet.)
Colorblind drivers have the benefit of a human brain, which is still orders of magnitude more powerful than Tesla's puny little GPU and which has been "trained" by millions of years of evolution. Colorblind drivers can also still clearly perceive whether a bulb is one or off and which bulb is on, which is harder than you'd think in various lighting conditions. It turns out that the human visual system is pretty darn amazing really.
Last edited:
I've known such people. It's a shitshow. Mostly just guess and look at other drivers for clues. And they're terrified of driving into a place which has horizontal lights.
akraut
Member
There's an interesting whitepaper of unknown date from some of the early waymo mapping of traffic lights: https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/37259.pdf
I wonder, how much first-hand experience do you have with features that Tesla has "released" to the fleet? I would not call any of them "done" at the time they go to wide release.
You don't say.
Seems reasonable given Elon's track record with tweets.
You should perhaps spend less time getting excited by what Elon said and your assumptions thereupon and more time driving what Tesla has actually produced.
Somebody move this post to snarkiness.
strangecosmos
Non-Member
Have you seen Canadian traffic lights?
Are Canadian traffic lights actually different from the U.S. or Europe? I thought it was an international standard.
loooooool
Last edited:
Trent has displayed that his only purpose in this forum is to cheer lead Tesla and be pessimistic about anything non Tesla.
The ones I saw are different. the horizontal ones had red on both sides for one. Search for "Canadian traffic lights" on google images.
Also they are certainly not standard and on top of that US is not a signatory to the international convention that standardizes this this.
strangecosmos
Non-Member
I live in Canada and can attest that virtually all the traffic lights I’ve ever seen are the standard vertical ones you would be familiar with in the US. I’m not sure if I’ve ever seen those horizontal traffic lights before, although they apparently exist.
Apparently Texas has them too! How weird.
Europeans switch from red to red&yellow then green. Some of them have countdowns. The main difference is that the light is in front of the crossing, not behind.
They also have a single flashing yellow light.
strangecosmos
Non-Member
Really interesting paper out from Waymo today about using imitation learning to train a self-driving car to do path planning like a human driver.
Waymo blog post | Paper | Discussion on Gradient Descent (no trolling)
Waymo trained its network on ~50,000 miles of human driving. It makes me wonder what you could do with billions of miles of human driving.
Waymo suggests using imitation learning to make naturalistically behaving simulated vehicles, which can then be used for reinforcement learning. This is an exciting idea.
I also wonder if, once imitation learning and reinforcement learning in simulation has taken you to a certain point, it would then be productive to do reinforcement learning with the real world fleet. Disengagements, aborts, and crashes would be logged and uploaded, and used as punishments. The reward function might be miles between punishments.
Waymo blog post | Paper | Discussion on Gradient Descent (no trolling)
Waymo trained its network on ~50,000 miles of human driving. It makes me wonder what you could do with billions of miles of human driving.
Waymo suggests using imitation learning to make naturalistically behaving simulated vehicles, which can then be used for reinforcement learning. This is an exciting idea.
I also wonder if, once imitation learning and reinforcement learning in simulation has taken you to a certain point, it would then be productive to do reinforcement learning with the real world fleet. Disengagements, aborts, and crashes would be logged and uploaded, and used as punishments. The reward function might be miles between punishments.
Last edited:
and then there are traffic lights without yellow, there are "paper arrows like this:
LED traffic lights actually flicker noticeably (see on my videos) when viewed at 36fps.
And then all sorts of modifiers (and now we only work for bicycles)... Traffic lights are quite a messy business. And it does not help that different countries have different rules on how to interpret the signals. Some places interpret the arrow as "you have the right of way", some others it means "you do NOT have the right of way, just let all others pass, then go", ...
Similar threads
- Article
