Public perception doesn't matter that much because the customers are car makers. What matters is their technology, which seems to be state-of-the-art and certainly better than what Tesla has.
Is it though? What Google has is certainly better than the publicly available lane and crash assist Tesla has, but Tesla has already shown demos of true autonomous driving, in fog and noting pedestrians, etc. It likely just needs a lot of training but cars are already on the road with hardware.
Is Google ahead for a final product? Quite possibly, but they don't have a product and Tesla doesn't look far behind sortware wise, though it is hard to tell.
Krafcik mentions in his presentation that their cars are down to .2 disengages every 1000 miles (which would more aptly be described as 1 disengagement every 5000 miles), and their cars are actively seeking out challenging situations, it's not just easy/random driving. This is incredible. They are leagues ahead of the nearest competition.
Krafcik also mentions they are making the lions share of their progress now using their simulation engine, where they can model every conceivable variation of any bizarre/difficult encounter they have on real roads. In sim they accumulated a billion miles in 2016.
The popular idea that Tesla will just keep stuffing matrices down the throat of their training pipeline until a self-driving inference model emerges from the other end doesn't make a lot of sense.
Neural nets can take you most of the way there, but pattern recognition alone will not solve the driving problem to completion.
Waymo's autonomous platform is a frankenstein of various machine learning techniques, and much of it isn't glamorous, it's less contigent on big breakthroughs than it is on elbow grease. Google demoed as proof-of concept full autonomy in 2012, and much of what they've been doing in the 4 years between then and now is the tedious job of addressing and validating their system across the full spectrum of edge cases that must be dealt with if they ever hope to foist their safety critical software upon the public.
It's not clear to me that Tesla's current development paradigm will ever be sufficient to completely take the human out of the loop. Tesla's approach is incremental, and I suspect they'll have to make some big changes if they wish to fully close the gap. Waymo has kept their eye on the prize from day 1.
Also, how much of that data can you realistically send back to yourself? With Waymo they can pull the arrays from the cars nightly if they need to. Tesla, on the other hand, has to send the data back over a customers network connection which would be much more limiting. So Tesla might be getting tons of miles but if that's a billion not very detailed miles, it might not be as useful as a million incredibly detailed miles.
I think Tesla pays for the network connection to all of its cars. This sends back remote telemetry and other self-driving data. It also serves up software updates and monitors the cars network for intrusions via a VPN.
That network connection is using cellular, so at most it's an LTE connection but if I remember right, is actually just 3G. I can't believe they are pushing the kind of data people are talking about over that kind of connection.
It makes sense because virtually all machine-learning algorithms work better, and can learn faster, if you have more data.
Think of it this way: if Tesla wants to test a particular algorithm for a particular driving situation, they can "play it back" over an enormous amount of real-world situations. They will have tons more potential edge cases with which they can validate their algorithms.
Where exactly is all this data being stored and transferred? Nowhere. The car doesn't have the storage and it doesn't have the bandwidth and all indications are it isn't transferring anything that amounts to actual images of the many many cameras it has.
Sure, they can push a beta algorithm to cars and record high-level decision making between human & algo, verifying it's not totally out of whack. But that's hardly something that is going as training data into the models.
This used to be true. Modern machine learning needs way less data. The classic example is taking images and then transforming them in hundreds of ways (scaling, rotation, skew, etc) for training.
Big data is no where near as much a competitive advantage as it was three years ago. It seems not everyone outside the field has noticed that though.
I wonder if this would be true in the case of self-driving car algorithms, though (which I know nothing about). It always seemed like the hard part about self-driving cars was the 0.1% edge cases where something out of the ordinary could result in a catastrophe if not handled correctly.
Image classification seems like it would be very different, most importantly that 99.9% "correct" would be a great achievement, but for self-driving cars a .1% failure rate would be completely unacceptable.
Do you have more examples of "big data is not as much a competitive advantage", in the form of articles or research? I'm not in this field, but it's a fascinating development. It would be interesting to see to which degree it helps to perform automated transformations to increase the value of each piece of training data.
You need to think this through more carefully. What would "playing it back" give you? Presumably all that would give you are a bunch of incidents where the human driver disagreed with the algorithm's output. That's it. We don't know whether the algorithm is actually wrong. Maybe the driver made a mistake. Maybe the driver wanted to make an illegal turn (very common). Maybe the driver was lazy and made a rolling stop. A human still needs to go through each incident that is flagged and manually label it.
It makes me wonder, though, if Google is currently at a disadvantage because their fleet of self driving cars is teeny compared to what Tesla has and will soon have. That is, Tesla has tens of thousands of cars with Auto Pilot sensors on the road, and (afaik) they have access to all of that real-world data to train their algorithms.
A big advantage I see in Tesla's strategy is that all of their cars are now shipping with full self-driving hardware. Even if that hardware isn't actually used to control the car, Tesla will have an order of magnitude more real-world data than anyone else.
But can they actually send all that data back to their data centers? If they are really capturing that much data, they would need to send it back using consumer home internet connections which doesn't seem realistic. I don't think the fleet size is as large an advantage if you can't actually get 100% of the sensor data.
It's important to note Google/Waymo's tech is limited to places that have been heavily manually mapped and test driven. They're basically creating a virtual track for their cars to follow, and presuming that changes in the real world can be detected and distributed to the entire fleet.
Other companies seem to be taking a generalized AI approach. Their cars work everywhere, but how well?
It'll be exciting to see which approach works best.
This information is not current, but I know that they used to detect traffic lights by looking for them in their maps. By that I mean they knew where the traffic light should be, then did the basic geometry to figure out where that is in their image, and then determined the color of the pixels to figure out what the light was saying. That absolutely would not work without the maps. I am willing to bet that their approach fundamentally depends on the maps. That's fine, it just means they can't work in areas that aren't already mapped (and you have to have one of the maps stored locally in the car, as those files are gigantic).
> they knew where the traffic light should be, then did the basic geometry to figure out where that is in their image, and then determined the color of the pixels to figure out what the light was saying.
This has to be really old information, though. They have that video of the car detecting school bus stop signs and a police officer directing traffic. A stop light is child's play after that.
Perhaps if the tech is good enough for now, it will become somewhat widespread, and then force a de-facto redesign of the traffic "API." E.g. maybe lights will have hidden IR identification beacons intended for self driving cars. Maybe traffic police will have special manual beacons that work on those cars while directing traffic, etc.
A software driving in a truck standing across the road, because confusing the truck is a road sign is nothing I would use autonomously. [1]
And I would not call it autopilot, especially being in a product.
I broadly agree that Tesla may take the lead. That said, I know they've announced autopilot being capable of handling fog and various weather conditions and have separately demoed their full autonomy, but I haven't seen anything combining the two.
Public perception doesn't matter that much because the customers are car makers.
I'll bet the number of people that buy a individual Intel CPU are a fraction of a percentage of the general population. So why do I see Intel ads on television?
Is there evidence that Intel Inside had a margin gain for the CPU product? Although not the tone of this article, the graphics suggest the II branding campaign had so much less to do with their growth than general computing trends.[0]
To answer your Intel/TV ad question more directly: because there are a lot of people on Intel's payroll whose salary directly depend on them placing ads on TV.
I do not demand you see my perspective, not even sure that's how I see it. But really, how do we know that Intel's branding campaign was valuable to them? It seems obvious, but if it's super obvious, it should be easy to explain.
Because people buy pre-assembled computers and Intel with its "Intel inside" had made it so that consumers will pay more if Intel is inside. Not sure if people will pay more for the same car with Google's sensor.
Public perception is irrelevant because self driving cars right now are irrelevant.
It's very much a binary situation like Siri. Either the technology works and is 100% effective under all conditions or it doesn't. If it doesn't then people aren't going to look at it as a key purchasing differentiator because they won't really rely on it. Because of this I think we have at least a decade if not more before it matters who is winning the race or not.
Not to take this into 'silly analogy' territory, but I have many knives that dull after two weeks of use and others that stay sharp much longer; some that work great to fillet fish and some that I can pry a screw out of an oak log with anf be fine; but none that can do both.
And I have blankets that I sleep great under in winter and summer and others that make me wake up in a puddle of sweat. And some I take camping and others that are very comfortable but very delicate.
So no I wouldn't say that anything I own works 100% of the time. Some work more than others, and good stuff generally works in several situations (or not, when it's optimized for one use case) - but overall, if your threshold for success is 'pretty well', then any self driving tech that isn't worse than humans would be good enough.
Is Google ahead for a final product? Quite possibly, but they don't have a product and Tesla doesn't look far behind sortware wise, though it is hard to tell.