[IAmWorried]

Wow. I had no idea that these things had such fundamental limitations. I'd feel a lot less safer using one now, to be honest. If you had to estimate, how long would it be before AVs would be able to do all their calculations locally? Is that something that's even under consideration at the time?

[ricardobayes]

It's a well-known thing in AV circles that some companies were actually faking their route by hardcoding them for demos etc. Also one shortcut is to use HD map data, and relying on 100k+ LIDAR sensor data that will never will be cost-effective. I know of only a handful of companies that rely on 'cheap' sensor data (cameras/radar) and use real-time calculations.

[michaelt]

> relying on 100k+ LIDAR sensor data that will never will be cost-effective

Eh, the $100k prices are mostly for high-end, low-production-volume research LIDARs.

There are LIDAR manufacturers claiming they can achieve prices of $500 [1] or even $200 [2] if they were fitted to every car made.

[1] https://guidehouseinsights.com/news-and-views/lower-cost-lid... [2] https://www.opsys-tech.com/post/are-we-nearing-a-shakeout-in...

[xyzzy123]

If a full time driver earns 25k-50k that's sort of equivalent to every regular taxi and uber having a $500k / $1M driving package onboard already (at 4%), right?

I don't know the industry but very naively it doesn't seem to me like $100k really breaks the unit economics? Particularly if volume / learning curve can reduce the price?

[ricardobayes]

You are right in the way that the industry seems to be shifting from general-purpose self-driving to transportation/hauling.

[papito]

Well, THAT has a Dropout-like HULU miniseries written all over it.

[marvin]

Tesla’s AV design philosophy differs from everyone else’s for just this reason. The idea of using centimeter-accurate models of the environment and having an always-on network connection is fundamentally at odds with handling surprises.

I don’t think a current Tesla would have a good time with no GPS either, but philosophically they stand to better handle intermittent unreliability and adverse conditions once they figure things out.

[ls15]

It is probably a chicken-and-egg problem. They need the data from these scenarios, so they can train the local model for later iterations accordingly.