Tesla’s autopilot won’t achieve full autonomy, Waymo CEO says

Lidar is almost like adding a hand crafted dimension to the decision making process. It makes the system more explainable, but doesn't necessarily mean it will help achieve better results. I look at ML in games like chess and go. Originally, they had a set of hand crafted features, like the concept of "eyes" in go. But later iterations removed those and the network performed better. I think lidar is the same thing where it'll add unnecessary complexity and ultimately hurt in learning.

Paragraphs 1-3 discuss legacy automakers. Only paragraph 4 references Waymo.

> You talk about OTA updates as if that’s some huge technological moat?

Of course they're not a technological moat. These auto makers are hundreds of thousands of employees. They're likely an organizational nightmare and stuck to the way things have been done. The fact that they don't have over the air updates or have to iterate on an annual cadence is evidence of something deeply flawed in their organizational structure. And I'm not confident that they can overcome it.

I think OTAs are a moat but highlight an important incentive difference. Traditional automakers want to sell you next year's model of the car. Tesla seems like they want your car to get better over time.

Perhaps it is organizational cruft or the dealer network but consider even the simple case of adaptive cruise control systems in cars today that have radar + cameras. Most of those systems bounce between lanes and can't even stay centered well. This is something that could be a software fix but you'll have to wait for the next N models when they redesign the car to see it.

Infotainment systems are another example - they are almost universally crappy and it is rare to get updates OTA. Shipping a car with capable hardware and slowly improving software capability imo is a smarter long-term approach.

Time will tell though.

One is depth-from-defocus combined with "overlap" hierarchy information that can be used to build a depth-map. This is the technique Tesla talked about. Interestingly, instead of manually coding up this algorithm, they simply trained a neural net to do it! The trick was to use the radar data as the "ground truth", and train the NN to predict it based on the optical input. It works quite well, and there are videos floating about showing how a Tesla can build up an accurate 3D model of the world around it as it drives along.

There are other approaches as well. For example, the way birds gather depth information is they move their head. One eye takes two "snapshots" of the scene in rapid succession from slightly different positions, producing a stereo pair. A moving car is constantly shifting the camera positions. Note that you don't have to be moving the cameras side-to-side, moving along the road can be enough, as the shift of the focal plane can be used as an additional channel of information for the depth-from-defocus method mentioned above. This is what some mirror-less cameras do: they shift focus back-and-forth a bit to gain more information about the scene. A moving car has the same effect.

Tesla has recently implemented some variant of this latter technique, they also talked about this publicly. Essentially, the first version of the self-driving hardware didn't have the computing power to analyse multiple frames together like this. The current version has the compute power to include the last few frames as well as inputs to the NN, which gives it not just more temporal stability but also the opportunity to extract depth information from motion. (However, I'm not sure if they've actually used motion in this specific way.)

Elon said coronavirus would be finished in the US by April, 2020. How'd that work out for him?