Hacker News Reader

Comment by puranjay | Hacker News Reader

I'm curious where Tesla fits into this scheme. In terms of road miles, I reckon Tesla would have Waymo beat easily. They have far too many cars on the road and thus, way more data.

Does this factor into the quality of self-driving tech at all?

View on HN

It's not that simple though. What so you consider "data" when it comes to self driving cars? Waymo collects high resolution data with every car (multiple GB or possibly TB of data per car per day).

Tesla hasn't sent very high resolution data back home yet. Though they are looking to get video from customers now to help make this go faster[0].

So while Tesla has more vehicles on the road, that doesn't mean they are getting better data for self driving car research.

[0] https://electrek.co/2017/05/06/tesla-data-sharing-policy-col...

They may not be sending data back home via the mobile network, but you can bet your boots they make exhaustive logs of Autopilot disengagements and store them locally to be siphoned off at the next service.

I wonder if they have WiFi base stations collecting data at Supercharger points? That's what I'd do.

Why not use owner's home network?

The logs might be pretty big, so it might have a significant impact on people on metered internet connections.

Not if it's only greyscale camera data and some point cloud data which is heavily compressible.

But then you're back to square one : that data is much less detailed than what waymo gets.

Machine vision is usually an exercise in data reduction. Greyscale is typically used in most computer vision applications, so it may be the case that Waymo has more information but that is not necessarily of any benefit.

The video has constant motion though, which makes compression much less effective.

Do owners generally connect their cars to their home wifi? (I don't know, is this a thing with Teslas?)

Even if so, uploading gigabytes of data every day is going to be noticed. While I'm sure they have the legal right to do so (via their TOS / EULA) they probably still don't want to emphasize that they're gathering that much data (unless you opt in to their fleet learning stuff).

>Do owners generally connect their cars to their home wifi?

Generally I think they do to access the Tesla App features. Tesla also have a network connection so they may already be uploading the information anyway over cell networks.

Tesla's don't connect to LTE networks by default?

If not, I find it shortsighted and rather cheap that each car doesn't have it's own $50/month data plan (or whatever) to phone home important data.

At 80k new cars a year that's a 4mil a month waste of money. Tesla couldn't survive giving each car a 50/mo data plan forever.

Waste of money?

To have a constant high speed IoT-like connection to the web for real-time updating and such?

I am sure some sort of volume discounts could be negotiated as well.

Teslas do connect to LTE. Why would they pay that much for a data plan? In the US, you can get mobile hotspot plans for $3/month + $10/gbyte. That's the consumer price; wholesale over 100s of thousands of cars has got to be cheaper.

I definitely wouldn't be happy about Tesla uploading gigabytes of data on my metered connection. Even if it was unmetered, most people don't have very fast up-speeds and that could take days to upload and heavily impact their speeds. Absolutely unacceptable.

What if they had agreement with xfinity/comcast, which opens up a separate wifi hotspot anyway?

> exhaustive logs of Autopilot disengagements

That would still be only one slice of the many many situations you need to master, and that Waymo already has. E.g. something as simple as entering and exiting a small roundabout, like that YouTube video showed a Tesla couldn't do. Miles driven by Tesla owners aren't purposefully looking for those situations when they engage autopilot.

Are you sure? Again this is just how I'd do things, but after I had a bunch of disengagement snapshots, the next thing I'd do is start taking snapshots of similar situations. We live in an age of easy and super-accurate image classification, it wouldn't be too hard to start a snapshot every time the front camera view contains a roundabout.

What I'm trying to say is that the number of cars on the road and miles driven aren't what matter to their autopilot learning, when compared to other self-driving companies. Most people use their car to go places comfortably, not to look out for learning experiences for Tesla.

With how long commutes are in the US, a very very small fraction of miles driven by people on autopilot will be other than "the car was already on the highway, autopilot changed lanes a few times, and after a while was manually disengaged to take an exit".

That small fraction is what could be compared with the miles and number of cars driven by the other companies, because in the latter case it's hired drivers actively looking for conditions of interest to the engineers.

As of 2013 Google's self-driving cars collected about 1 GB per second of data [1], today it must be even more. There is no way Telsa can upload nearly as much data from customer's cars. I also belief that their team is smaller and they started much later.

[1] http://www.kurzweilai.net/googles-self-driving-car-gathers-n...

You shouldn't need to upload the entire 1gb. The central AI can query the cars to get just the bits it has questions about. Part of the AI's job is learning which data to care about and which data to ignore.

The fact that the trees got new leaves again this year and it's dark at 2am and the sign at the Chinese Restaurant is still red don't to be relayed.

More data isn't automatically better. The hard part isn't getting some magical quantity of data the hard part is asking the right questions of it.

A self driving system needs to solve this problem (the frame problem) anyway. So you get intelligence compression for "free".

In a sense, intelligence is compression.

You have essentially just described distributed learning. Only transmit the gradients back to the central server, not the data.

That'd be one way of doing it. You could also just send the cars experimental queries to upload, and still do the leaning centrally.

This is naive. You don't know what data you will care about in a year. The ability to have and keep all the data is incredibly valuable.

Historic data is certainly usable. I'm not sure where your idea comes from that it often provides value far beyond what's in the short term cache.

What makes you believe that?

How many things happened today that won't happen next week, somewhere?

And please don't call me naive. That's rude. Disagree substantively; use your words.

Why do so many on HN feel the need to structure their responses like this:

> Something Rude. Real argument, which is stronger with first sentence removed.

You may be right (or wrong, I don't know) but opening with rudeness or insults makes it more difficult to have meaningful conversations.

Rudeness? Insults? I said "this is naive", not "you are naive."

Surely we're all adults and can talk about ideas candidly without having to beat around the bush and tell each other little white lies.

A while ago you told someone, "You speak in broad generalizations that miss many points and make many faulty assumptions. "

Any difference between your statement and mine comes down to how individuals respond to specific words. How about we rise above such pettiness and focus on the subject matter?

You've added nothing to this discussion and have wasted people's time.

> You've added nothing to this discussion and have wasted people's time.

What is your response doing?

Very first sentence of your source says "Google’s self-driving car gathers 750 megabytes of sensor data per SECOND!", here goes one fourth of the alleged 1GB.

Then this is about sensor data or raw data, but there is thing known as compression which is quite common and can shrink size by an order of magnitude.

75MB/s seems closer to reality and manageable.

Wow. I reckon this must be even more now.

Google is also a better computing company than Tesla, I imagine, whose strength is hardware, not software.

Last week Tesla began collecting short sections of video data from those with their new in-house Autopilot hardware. So really, Tesla's just getting started, and they're doing it with underpowered sensors and compute.

https://www.theverge.com/2017/5/8/15581094/tesla-autopilot-u...

Waymo was ahead of where Tesla is now back in 2012. In a demonstration capacity, Waymo's cars were working 5 years ago, but in the intermittent time Waymo has been validating their software against every possible edge case they might encounter on a real world situation, and Tesla will have to do the same one way or another before their autonomous OS can operate safely without a babysitter.

It is a matter of point of view, how many waymo car have been sold to customers ? how many waymo cars are out of the waymo car factories every months ?

There are several things to tackle and different approaches to things, the DARPA grand challenge is the practical one and waymo did not took part in any edition AFAIK.

From Waymo's wiki page: "Google's self driving car project was formerly led by Sebastian Thrun (...). Thrun's team at Stanford created the robotic vehicle Stanley which won the 2005 DARPA Grand Challenge and its US$2 million prize from the United States Department of Defense. The team developing the system consisted of 15 engineers working for Google, including Chris Urmson, Mike Montemerlo, and Anthony Levandowski who had worked on the DARPA Grand and Urban Challenges."

How many people have died in Waymo cars?

The miles aren't equal, though. Waymo's cars have an amazing array of sensors on them, including 360 degree LiDAR, which Teslas don't have.

teslas don't have lidar, so it's hard to say how useful their data will be in the long term

True, but if it turns out LIDAR is overkill and that self-driving can be done with only cheaper sensors (cameras), Waymo might be behind the curve on the use-case that the market will opt for.

When it comes to your safety, would you rather overkill or underkill? Especially when it comes to something as important as self-driving, I think it's absolutely reckless to try to "maximize profit" and "get it out there ASAP".

It would be an absolutely disaster if the technology as a whole gets a horrible reputation, or even worse, banned in places, just because some company skipped some steps and wanted to be the first or cheapest to the market.

I'm sure that Google gathers so much data from human drivers using Google Maps and Waze that they can build a decent model.

Why do you think Google maps data would be as useful as the camera and steering input data Tesla can collect?

Why do you think Tesla is able to collect all or any of the camera and steering input data? And what proof do you have that the little they are able to get off is actually at all valuable.

Ok, fine, we don't know what either company is collecting, but google maps can ONLY collect GPS and speed data, whereas Tesla has the POTENTIAL to collect gps, speed, camera, steering input, braking input, and acceleration input.

Yes and no. In the bay area and other big cities Teslas are definitely there but there are a lot of other areas in the country where Waymo probably spent the money to drive and collect data. So we probably will not know the reality.

Reply on news.ycombinator.com