Hacker News Reader

Comment by paulsutter | Hacker News Reader

> We expect that worldwide regulatory approval will require something on the order of 6 billion miles (10 billion km). Current fleet learning is happening at just over 3 million miles (5 million km) per day.

This seems very significant for Tesla vs competitors. Yes Google has a strong technology lead today, but how long will that last when Tesla is collecting more miles of data every day than Google has collected in 5 years? (Sincere question) Not to mention Apple and existing car vendors, who each have 0 million miles of experience.

Tesla should reach 6 billion miles very quickly once the model 3 is out.

View on HN

It seems to me the data tesla are collecting are not unique. How many times logging the same driver driving the same 20 mile commute before you're in diminishing returns?

With deliberately chosen conditions, scenarios, and routes, it seems to me Google could be collecting data that are just as useful although maybe covering much a smaller number of miles.

I don't know where you live, but my daily commute is never the same. They either start/end at different times, or I have to take different routes, or the weather is different. I think it would take a lot longer than 3 years before they start seeing actual diminishing returns on mileage collection.

More than 3 years? I would confidently wager that noticeable diminishing returns would occur after 3 days, and very significant diminishing returns after 3 weeks.

Think about seasons, varying weather, light levels, traffic fluctuation, construction, changing parking patterns, delivery drivers/stopped trucks...the list goes on. It would be very useful to have that occur within a relatively known route, where you can isolate and model what isn't understood already.

I can't imagine that the data from carefully selected scenarios would be anywhere near as valuable as the data from random, unpredictable, real world driving.

What does that data give you? Raw sensor data + steering angle + acceleration/braking is not particularly useful in of itself. Nobody is trying to shove large amounts of data through a blackbox model and directly trying to predict car controls from images. This is not MobileEye's approach and almost certainly isn't Tesla's approach. The bottleneck is data that is decomposed and labeled in greater detail (lane markings, obstacles, bicyclist hand signs, etc) - and you need human annotators for that.

Geohot is pretty much claiming to do that!

http://www.theverge.com/2016/6/6/11866868/comma-ai-george-ho...

Sure, some people are exploring the approach. This does not mean it works.

I recommend watching this [0] ted talk on the subject. In it, you'll see that Google, last year, already had solved bicyclist hand signs, lane markings, unforeseen obstacles, etc. And they did it __without__ human annotators. Just some straight up convolution neural networks.

[0]: https://www.ted.com/talks/chris_urmson_how_a_driverless_car_...

You're mistaken. Convolutional neural networks require annotated data. It's a TED talk; they're obviously not going to mention the very boring details of their annotation protocols.

Some quick napkin math: 3 million miles a day gets you 6 billion miles in 5 and a half years (not counting the ~100mm miles they already have). Since the daily mileage rate will only be growing exponentially, it's not unreasonable to assume that Tesla hit this number in under 3 years.

Not sure how they came up with 6 billion as the magic number, but if that's accurate, that's very soon. It also helps that there are no real government stakeholders in preventing the proliferation of self-driving cars. People just need to generally believe that they are much safer than human drivers, which of course is not true yet.

How much data has Google and Apple collected using LIDAR and such on their mapping/street view vehicles?

It's in the millions. Tesla's is in the hundreds of millions/close to a billion.

Technically, it's 0, because Tesla does not use LIDAR at all.

Almost every other car manufacturer is working on autonomous vehicles. They will eventually start to contribute to the number as well.

Great metric to establish that seems reasonable and is heavily skewed in their favour.

how many miles of data has googled accumulated is a unanswered question. Also does that mean 6B miles combining all autonomous vehicle manufacturers?

Google actually publishes stats on their self-driving cars. They're currently just shy of three million miles in the last seven years, only some of which are from autonomous mode: https://static.googleusercontent.com/media/www.google.com/en...

Reply on news.ycombinator.com