When robocars must be perfect, and when they need not be


A recent Waymo tester has been challenging Waymo cars to pick him up in unusual pickup spots. Some of the times, the problem is probably being solved by a remote human operator giving advice to the car. What many do not understand is that this is not a flaw, but probably the simplest and cheapest way to solve the problem.

Read more at When robocars must be perfect, and when they need not be


IBM prides itself on simulating the brain of a cockroach, and then a cat. I wonder if a simulation of a horse is not what one wants for a robot vehicle. A cat is not cooperative enough, a dog is too distractable, but a good simulation of the mind of a risk-avers grazer sounds like what I would like driving my car for me.

(Hi Brad, long time since Waterloo CSC)

And what's "the problem" that these companies are trying to solve?

I suspect that Waymo and Tesla are trying to solve a different problem.

Well, Waymo strictly is aimed at Robotaxi. Tesla also wants to do that, but in addition wants to sell cars to customers -- and sees that as the path to the robotaxi. I don't know if Tesla has, or needs to have an opinion on which one is better.

Tesla feels that it's a waste of time to make a car that can't go everywhere. If they can make one fast, they are right. If they don't get the breakthrough they want then they are probably wrong. If they get it, but get it late, that's still good for them, although once anybody gets it, the other top-notch teams will certainly do it in much less time.

Particularly Waymo. I mean, they get the resources of Google, which employs both the creator of deep learning, some of the best minds in AI, and the DeepMind team which are the world leaders in reinforcement learning techniques by a pretty large margin. I would expect them to be able to duplicate what somebody else does in enough time to continue the fight. They are also the masters at big scale data, beyond anybody else in the world.

Waymo is trying to build a better taxi service. Tesla is trying to build a robocar.

Tesla has made a car that can go anywhere. They're working on improving that car so it requires less human monitoring/input. Eventually I expect they'll allow that human input to be provided remotely, probably by tesla employees, at least in places with good enough reliable network connectivity. People with Tesla cars will have the best of both worlds. Passenger-only (or cargo-only) where available, ADAS FSD autopilot everywhere else. For long trips to a destination where passenger-only is unavailable, you can have your car wake you up when you get a few miles away. If you don't indicate that you're ready to take over, your car parks for you somewhere safe. A car should be able to go anywhere. Tesla's cars can.

Waymo hasn't made a car at all, and the taxi service they've made can't go very many places at all, and offers little benefit over Uber/Lyft (maybe a moderate benefit in the age of COVID). They hope to lower costs drastically, and maybe they will eventually, but it looks like it's going to take them a long long time to do so.

It's not clear to me what the problem is with Waymo. Is it really that hard for them to expand to more places?

Maybe it's primarily a regulatory problem...

Hopefully it's just the pandemic, and they'll be ready to roll out to dozens of cities shortly after the pandemic lifts. But then it's still a matter of figuring out how to drastically cut costs. I think having one employee for every 100 cars is a pipe dream, at least for the foreseeable future. And there are tons of infrastructure costs they'll have, both up-front and ongoing.

While in the past Waymo was called the reckless one (for testing on public streets without telling anybody) they are actually much more cautious than Tesla. That's part of why they have not deployed when Tesla would have in the same situation.

Tesla has not build a robocar that can go anywhere. They have built a driver assist system that can go anywhere. Huge difference. To put it in the terms of the former chief engineer on Autopilot, "You can't build a ladder to the moon."

Waymo's service today is cheaper than Uber, but only modestly. Of course on the one hand it really costs much more than that to run (as a prototype) but on the other hand at scale it will cost much less when you don't have to factor in huge NRE. There is no reason to make it that cheap yet, as they could not meet the demand.

It looks like the city of SF is their next target (after Silicon Valley.)

Tesla has not build a robocar that can go anywhere.

I did not say they did.

at scale it will cost much less

they could not meet the demand

At most only one of those two statements can be correct.

Add new comment