One interesting thing in the Swiss RE + Waymo paper is that the Testing Operations results (with a safety driver) are still better than the Rider Only results (with no safety driver), particularly for lower-speed (property damage only) collisions. This likely indicates that in the situations where a human decides to take over, they then outperform the AI (just as a human would have taken over and done much better in the emergency situation described in the post). It's not obvious how to leverage that to do better, but it is interesting.
> So while 60 altercations with the SFFD might seem like a lot, it works out to fewer than 1 incident for every 60,000 miles of driving.
If a person-year of driving is ~10k miles, this is one incident per 6 person-years. I think that’s actually kind of a lot!
If I talked to ten drivers and asked about their past 10 years of driving, I’d expect to hear about maybe one such incident. By that rough guesswork, the driverless cars are something like 20x as likely as humans to have such incidents.
That could be worth the tradeoffs for fewer high speed accidents! But, it seems like a bad failure mode.
Yes, I agree that it's a lot. I'm hopeful that they'll improve pretty rapidly since they've only really been dealing with these kinds of problems in the last year or two.
We are in the early days of autonomous vehicles with much to still improve on. But advancement is going to be relatively quick and the benefits to society huge. Have patience grasshoppers!
Once autonomous cars are perfected, personal ownership of automobiles will disappear. For the vast majority of people, cars sit around unused for much of the day. Cars are an expensive and depreciating asset that will almost never recover the cost to buy and run them. So why buy one if you don't have to?
In the not distant future, especially in cities and satellite suburbs, people will call for a car from fleets whenever they need to go anywhere. The wait will not be any longer than 10 minutes. Some portion of the cars will be constantly circulating in local areas to be able to reach people quickly while the vast majority will be parked in depots on the outskirts of cities, out of the way of day-to-day business. Of course, algorithms will attempt to predict usage requirements and have cars staged nearby for quick access.
These cars will be networked and communicating constantly, allowing them to drive faster and safer than human driven vehicles.
People will not need car insurance any longer. They will not have monthly car/lease payments. They will not need to worry about car maintenance. There will no longer need to be parking garages and street parking will disappear. Your home garage can be turned into a storage area, as it is for many or an extra guest bedroom.
Car dealerships will disappear. The car mechanic business will disappear. Auto insurance brokers will disappear. Best of all, those dumb open road advertisements ubiquitous from car manufactures will disappear.
Finally, humans will not be allowed to drive cars on public streets.
I don't think so! Part both Waymo and Cruise's strategy is to allow its cars to "phone home" for help if it encounters situations it doesn't understand. So human intelligence effectively augments what the AI can figure out.
One interesting thing in the Swiss RE + Waymo paper is that the Testing Operations results (with a safety driver) are still better than the Rider Only results (with no safety driver), particularly for lower-speed (property damage only) collisions. This likely indicates that in the situations where a human decides to take over, they then outperform the AI (just as a human would have taken over and done much better in the emergency situation described in the post). It's not obvious how to leverage that to do better, but it is interesting.
Excellent reporting and analysis! Thank you so much for the effort you put into these in-depth articles.
Thanks for reading Kevin!
> So while 60 altercations with the SFFD might seem like a lot, it works out to fewer than 1 incident for every 60,000 miles of driving.
If a person-year of driving is ~10k miles, this is one incident per 6 person-years. I think that’s actually kind of a lot!
If I talked to ten drivers and asked about their past 10 years of driving, I’d expect to hear about maybe one such incident. By that rough guesswork, the driverless cars are something like 20x as likely as humans to have such incidents.
That could be worth the tradeoffs for fewer high speed accidents! But, it seems like a bad failure mode.
Yes, I agree that it's a lot. I'm hopeful that they'll improve pretty rapidly since they've only really been dealing with these kinds of problems in the last year or two.
what do you think about this video? https://www.youtube.com/watch?v=-MUEXGaxFDA
We are in the early days of autonomous vehicles with much to still improve on. But advancement is going to be relatively quick and the benefits to society huge. Have patience grasshoppers!
Once autonomous cars are perfected, personal ownership of automobiles will disappear. For the vast majority of people, cars sit around unused for much of the day. Cars are an expensive and depreciating asset that will almost never recover the cost to buy and run them. So why buy one if you don't have to?
In the not distant future, especially in cities and satellite suburbs, people will call for a car from fleets whenever they need to go anywhere. The wait will not be any longer than 10 minutes. Some portion of the cars will be constantly circulating in local areas to be able to reach people quickly while the vast majority will be parked in depots on the outskirts of cities, out of the way of day-to-day business. Of course, algorithms will attempt to predict usage requirements and have cars staged nearby for quick access.
These cars will be networked and communicating constantly, allowing them to drive faster and safer than human driven vehicles.
People will not need car insurance any longer. They will not have monthly car/lease payments. They will not need to worry about car maintenance. There will no longer need to be parking garages and street parking will disappear. Your home garage can be turned into a storage area, as it is for many or an extra guest bedroom.
Car dealerships will disappear. The car mechanic business will disappear. Auto insurance brokers will disappear. Best of all, those dumb open road advertisements ubiquitous from car manufactures will disappear.
Finally, humans will not be allowed to drive cars on public streets.
A Tesla stopped working and got stuck in the middle of a busy road for 9 hours because it couldn't be moved, report says
By Jyoti Mann Business Insider
September 3, 2023
https://www.businessinsider.com/tesla-stopped-working-stuck-busy-road-nine-hours-2023-9
Is AGI a necessary condition for driverless vehicles to be safely deployed?
I don't think so! Part both Waymo and Cruise's strategy is to allow its cars to "phone home" for help if it encounters situations it doesn't understand. So human intelligence effectively augments what the AI can figure out.