Could you say more about the "they had to send out a manual driver" bit?
Car encountered an unmarked dead end and got confused turning around (there was a badly-parked car on the narrow road). Car called for help and we were told a person was en route. A few minutes later, a guy arrived in a truck and manually drove us to our destination.
Or they can mark a vehicle as parked so the car won't wait for a stationary vehicle it thinks has right of way.
But they can't directly drive the vehicle. The whole software and Comms stack is far too slow to do that safely.
Pay 5G networks to have priority data feeds for video, and send all the data back to an operator with 50ms glass to glass latency. Then have the remote operator drive.
Always have the car in a position to take over and stop incase the comms link gets cut. Usually that will involve either slamming the brakes on hard or gently, depending if there is an obstacle ahead or someone following behind.
That means they no longer need to handle 99.9999999999% of cases, but instead just handle 99.9% and detect the rest for a human to do.
You are potentially overlooking application level latency, just because the packets arrive in 50ms time does not mean the video latency is upper bounded to that. I think 100-200ms is more realistic.
If your one way network latency is 50ms, it takes 150ms minimum to send a packet on 1.5x round trips which is necessary to confirm round trip connectivity and notify the other party. If you can assume clocks are synchronized you can send packets one way in 50ms to confirm connectivity and latency.
> Usually that will involve either slamming the brakes on hard or gently, depending if there is an obstacle ahead or someone following behind.
you’re specifically in a situation in which the car has failed, so im not sure how you can assume the car will make these kinds of decisions correctly.
> I personally think they should just update the stack so they can do realtime control.
Lets assume video latency of 100-200ms can be achieved, its still potentially dangerous. There are however some companies i am aware of in certain countries that are doing it anyway!
Even if glass to glass latency was 1ms, perhaps its not a technological issue but a PR issue. Maybe they don’t want news headlines about people driving cars with mario kart steering wheels or it’s easier to get regulatory approval by reassuring the public they will do the safest thing possibly.
I wonder what the air force does with remote drone pilots?
I think you may be an outlier and early adopter type person and may not represent the average customer.
Props to Waymo for what they've accomplished; they've been very clever in working around the limitations of the tech. But it sounds like it has a long way to go before they're ready for general-audience usage.
[1] E.g., the Crossing the Chasm model, although this is more the behavior of the "visionary" segment than what he calls "early adopter"; scroll down to the first big graph here: https://thinkinsights.net/strategy/crossing-the-chasm/
Yup! It as far more forgivable than an Uber driver taking the wrong turn half a dozen times in a row. It was half the cost, before tip. And it was up to four times as long a wait, not always four times as long. All of that more than makes up for the fun.
It’s not ready for mass roll-out. But there is a massive beachhead to grab.
Toyota lost a lot of money the first few years they sold the Prius, but they were happy to do it because they believed they'd work out the production cost issues over time. And they were right. Clearly Waymo's going for a similar strategy here. But one difference is that Toyota at that point had decades of experience wrangling the cost curve for cars. It'll be interesting to see if they can get capital cost plus mapping cost plus operational cost down low enough so that they can at least break even on a ride.
Only if they don't stop off at a tourist trinket store.
It's pretty clear that with Waymo the first response is to fail safe: when in doubt, pull over and ask for help. I was expecting that the help could include a semi-assisted remote override mode. The remote support would not be turning the wheel and pressing the gas, but giving it higher-level overrides, like "do a u-turn here". So in that case if it loses touch it's reasonable to resume traveling if it now can or once again pull over and ask for help.
So either you're no worse off, or the link is fine and you don't need to send someone out at all. Sounds pretty good to me!
Like I said originally, if it was safe for the car to be driving itself it already would be.