Autonomous Car Progress

[ThomasD]

You don't need 3rd party software for that!

The whole video is worth watching. 16:15 is terrifying.

It should be my Tesla almost hit a train

 

[Dewg]

The truth is that driverless vehicles need to be near 100% flawless. Human drivers achieve 1 death per 100 million miles of driving in the US.

Absolutely. I've seen that statistic before, 1.06 deaths per 100 million miles. I read an article that did some amazing statistical math, using a metric I've never heard of called "deaths per million vehicle-years", which they then calculated out based on the average driver driving 12,000 miles per year, and came up with human drivers achieve 1 death per 428 million miles.

I know Tesla releases safety reports quarterly, but I don't know if there is a regulatory reporting mechanism for Tesla to report the number of miles driven on AP/NoA/FSD. All I can find is what Tesla reports, which I know several people here will dismiss as lies. In Q2 last year, Tesla claims one crash for 4.41 million miles on AP, vs one crash for 1.2 million miles without AP.

NHTSA also released a report Early Estimate of Motor Vehicle Traffic Fatalities for the First Half (January-June) of 2021. In which they showed a sizable increase in miles driven (probably due to the slowing down of the pandemic) of about 173 billion miles, a 13% increase. Based on that report, the fatality rate for the first half of 2021 increased to 1.34 deaths per 100 million miles.

I was able to find some numbers on Tesla deaths on AP, and 2021 had 4 deaths claimed on AP, and 1 death verified on AP (it was a pedestrian that was struck by someone on AP while changing a tire). Using Tesla's safety reports for each quarter of 2021: Q1=4.19 million miles on AP, Q2=4.41 million miles, Q3=4.97 million miles, and Q4=4.31 million miles. That gives us a total of 17.88 million AP miles in 2021. The 1 verified death in 2021 on AP would then be 5.6 deaths per 100 million AP miles in 2021.

If I do the same math for 2020, there were 6 deaths claimed on AP, and 2 deaths verified on AP. Using Tesla's safety reports for 2020: Q1=4.68 million miles on AP, Q2=4.53 million miles, Q3=4.59 million miles, Q4=3.45 million miles. That gives us a total of 17.25 million AP files in 2020. The 2 verified deaths in 2020 on AP would then be 11.6 deaths per 100 million AP miles in 2020.

2019, there were 9 deaths claimed on AP, 7 deaths verified. 13.55 million miles drive on AP. That gives us 51.6 deaths per 100 million AP miles in 2019.

So far the pattern seems to be getting better as time goes on with Tesla AP.

Sources: Tesla Vehicle Safety Report and TeslaDeaths.com: Digital record of Tesla crashes resulting in death

 

[diplomat33]

XPeng released a demo to tease the engineering version of their L2 "door to door" system called City NGP.

https://twitter.com/x/status/1543837563819282432

I am not claiming XPeng's City NGP is better than Tesla's FSD. The video is edited and sped up. However, I think the video is informative because it shows some of the features that City NGP will have that are very similar to NOA and FSD Beta. I do like that the voice assistant communicates with the driver what the system is doing.

 

[Daniel in SD]

If I do the same math for 2020, there were 6 deaths claimed on AP, and 2 deaths verified on AP. Using Tesla's safety reports for 2020: Q1=4.68 million miles on AP, Q2=4.53 million miles, Q3=4.59 million miles, Q4=3.45 million miles. That gives us a total of 17.25 million AP files in 2020. The 2 verified deaths in 2020 on AP would then be 11.6 deaths per 100 million AP miles in 2020.

2019, there were 9 deaths claimed on AP, 7 deaths verified. 13.55 million miles drive on AP. That gives us 51.6 deaths per 100 million AP miles in 2019.

Haha, might want to check your math, if that were true Autopilot would have been banned long ago. Tesla's numbers are the rate of airbag or active restraint usage (seatbelt pretensioners) which they say correlates to collisions over 12mph.
My point is that anyone saying that FSD is close to human performance based on personal experience is wrong because it's literally impossible for an individual to determine that within their lifetime.

 

[Dewg]

Haha, might want to check your math, if that were true Autopilot would have been banned long ago. Tesla's numbers are the rate of airbag or active restraint usage (seatbelt pretensioners) which they say correlates to collisions over 12mph.
My point is that anyone saying that FSD is close to human performance based on personal experience is wrong because it's literally impossible for an individual to determine that within their lifetime.

Since Tesla doesn't give us the total number of AP miles driven in a given year, I had to extrapolate. The pattern would seem to hold though, less deaths per AP miles each year. And you're right, we're not near 1.06 per 100 million miles, but well on our way.

 

[Daniel in SD]

Since Tesla doesn't give us the total number of AP miles driven in a given year, I had to extrapolate. The pattern would seem to hold though, less deaths per AP miles each year. And you're right, we're not near 1.06 per 100 million miles, but well on our way.

Not sure exactly what you're saying. I'm sure the death rate for Tesla vehicles is better than 1.06 per 100 million miles either with or without AP on because of demographics and the fact they're much safer than the average car on the road (average car on the road is much older and doesn't have AEB, lane departure avoidance, etc.).
If you're talking about FSD without a safety driver then it seems like they've got a long way to go to get to 1 collision (>12mph) per 2 million miles (that's what Tesla claims human drivers of Teslas achieve). My guesstimate would be 1 per 2000 miles if I'm being an optimist, so 1/1000th of the way there!

 

[ThomasD]

I would be happy with self driving I 75 through Atlanta and the I- 75 I-24 corridor around and through Chattanooga TN

 

[AlanSubie4Life]

I had to extrapolate.

No this is not extrapolation. As you say you need to know the number of AP miles. Or the number of accidents.

They had 1 billion AP miles in total by 2018 (Source: Tesla). Likely considerably higher now! Just guesstimate 10 billion or whatever over 4 years and call it good. If you want to be optimistic call it 20 billion. Note: These are random numbers.

(Would imply 0.05 deaths per 100 million miles, which is probably within an order of magnitude. But as I said, just random made-up numbers. Useless.)

 

[Dewg]

Not sure exactly what you're saying. I'm sure the death rate for Tesla vehicles is better than 1.06 per 100 million miles either with or without AP on because of demographics and the fact they're much safer than the average car on the road (average car on the road is much older and doesn't have AEB, lane departure avoidance, etc.).
If you're talking about FSD without a safety driver then it seems like they've got a long way to go to get to 1 collision (>12mph) per 2 million miles (that's what Tesla claims human drivers of Teslas achieve). My guesstimate would be 1 per 2000 miles if I'm being an optimist, so 1/1000th of the way there!

You're probably right on AP miles - I found a site that did the calcs and estimates, but it was a few years ago, before the Y came out. He estimated over 5 billion AP miles around 2021. Given those numbers, yes Tesla cars on AP are safer than 1.06 deaths per 100 million miles.

Since there is no L4 or L5 Tesla yet, we don't have any numbers to estimate FSD without a driver. All we have are L2 numbers which have a human driver supervising. I can't find any deaths while on FSD Beta, but I can't guarantee that number. Since FSD Beta is limited to city streets where speeds are much lower than freeways/highways, it's reasonable to assume there wouldn't be that many deaths over 100 million miles of use.

The same can be said for Waymo and Cruise, we're likely to see very few deaths per 100 million miles due, in part, to the slower speeds in a collision, and the fact that current L4 and L5 vehicles on public roads are limited to city streets without a driver, and don't operate on freeways/highways.

 

[Daniel in SD]

You're probably right on AP miles - I found a site that did the calcs and estimates, but it was a few years ago, before the Y came out. He estimated over 5 billion AP miles around 2021. Given those numbers, yes Tesla cars on AP are safer than 1.06 deaths per 100 million miles.

Since there is no L4 or L5 Tesla yet, we don't have any numbers to estimate FSD without a driver. All we have are L2 numbers which have a human driver supervising. I can't find any deaths while on FSD Beta, but I can't guarantee that number. Since FSD Beta is limited to city streets where speeds are much lower than freeways/highways, it's reasonable to assume there wouldn't be that many deaths over 100 million miles of use.

The same can be said for Waymo and Cruise, we're likely to see very few deaths per 100 million miles due, in part, to the slower speeds in a collision, and the fact that current L4 and L5 vehicles on public roads are limited to city streets without a driver, and don't operate on freeways/highways.

Well VRUs (pedestrians, cyclists, etc.) make up more than a quarter of deaths and the collision rate is higher in cities so I'm not so sure that it's reasonable to assume that.
Obviously you can estimate how FSD would perform without a safety driver (what do you think Waymo and Cruise did? They didn't just remove the safety driver to find out!). An FSD Beta user can only get a very rough estimate since you don't have the simulation tools that Tesla has to run the counterfactual of what would have happened had you not disengaged. And of course that analysis is highly dependent on predictions of what other human drivers will do (especially with FSD Beta and its propensity to steer into oncoming traffic. How often will the other drivers successfully avoid you?). Obviously it's nowhere close to human averages but it's fun as a thought experiment to estimate. I think it's a severe collision every 200-2000 miles.

 

[Dewg]

Well VRUs (pedestrians, cyclists, etc.) make up more than a quarter of deaths and the collision rate is higher in cities so I'm not so sure that it's reasonable to assume that.
Obviously you can estimate how FSD would perform without a safety driver (what do you think Waymo and Cruise did? They didn't just remove the safety driver to find out!). An FSD Beta user can only get a very rough estimate since you don't have the simulation tools that Tesla has to run the counterfactual of what would have happened had you not disengaged. And of course that analysis is highly dependent on predictions of what other human drivers will do (especially with FSD Beta and its propensity to steer into oncoming traffic. How often will the other drivers successfully avoid you?). Obviously it's nowhere close to human averages but it's fun as a thought experiment to estimate. I think it's a severe collision every 200-2000 miles.

Has your Tesla on FSD Beta ever veered into oncoming traffic? Mine never has. The worst it's done is attempt to move on a red left turn arrow on this recent version. Or are we taking a few data points and extrapolating them to the entire fleet?

One of these days I'll have to come down to SD and see how bad your FSD experience is.

 

[Daniel in SD]

Has your Tesla on FSD Beta ever veered into oncoming traffic? Mine never has. The worst it's done is attempt to move on a red left turn arrow on this recent version.

If there were oncoming cars that is veering into oncoming traffic!
I actually don't have FSD Beta though I have a friend with it. I'm waiting for all you beta testers to get it to driverless or at least L3.
What's your estimate of the >12mph collision rate if unsupervised?

 

[diplomat33]

Driverless rides in Cruise:

Without Driver, But With Coyotes: First Ride in a Cruise Robotaxi Through San Francisco

Tuesday night was the night. I rode in one of Cruise’s fleet of driverless robotaxis in San Francisco for the first time. Thanks to an Austrian friend who already had access to the Cruise app…

thelastdriverlicenseholder.com

 

[Dewg]

Driverless rides in Cruise:

Without Driver, But With Coyotes: First Ride in a Cruise Robotaxi Through San Francisco

Tuesday night was the night. I rode in one of Cruise’s fleet of driverless robotaxis in San Francisco for the first time. Thanks to an Austrian friend who already had access to the Cruise app…

thelastdriverlicenseholder.com

At the 5:50 mark on the first video, it appears the cruise continues to slow down after the light turns green, as if there was a delay processing the light change. I wonder if that was for comfort to reduce g-forces for the passenger.

 

[Daniel in SD]

At the 5:50 mark on the first video, it appears the cruise continues to slow down after the light turns green, as if there was a delay processing the light change. I wonder if that was for comfort to reduce g-forces for the passenger.

I bet it's programmed not to go too quickly through lights that just turned green in order to avoid red light runners.

 

[diplomat33]

At the 5:50 mark on the first video, it appears the cruise continues to slow down after the light turns green, as if there was a delay processing the light change. I wonder if that was for comfort to reduce g-forces for the passenger.

Yes, I think so. Passenger comfort is important for robotaxis. I know companies like Waymo and Cruise have put some work into making sure that the ride is as smooth as possible for the passengers. Going from de-acceleration to sudden acceleration when the light changed from red to green would have made the ride a bit jerky for the passenger. So the cars slows down to smooth out the acceleration.

 

[diplomat33]

NHTSA will investigate the crash last month involving the Cruise AV and a Prius:

According to the report, filed by Cruise Vice President of Global Markets Todd Brugger, a Toyota Prius entered an intersection after traveling straight via a lane designated for turning. The Cruise vehicle was attempting to make a left-hand turn across several lanes of traffic and had stopped to allow the car to turn.

The Prius was traveling about 40 mph in a 20 mph speed zone when it struck the Cruise vehicle, according to the filing. The Cruise vehicle was in “autonomous mode” at the time of the crash. It’s unclear if a safety driver, employee or other passenger was in the car.

U.S. safety regulators to probe crash involving self-driving car from GM-backed Cruise

Federal vehicle safety regulators will investigate a crash last month in which a vehicle struck a self-driving car from General Motors-backed Cruise.

www.cnbc.com

 

[Doggydogworld]

NHTSA will investigate the crash last month involving the Cruise AV and a Prius:

U.S. safety regulators to probe crash involving self-driving car from GM-backed Cruise

Federal vehicle safety regulators will investigate a crash last month in which a vehicle struck a self-driving car from General Motors-backed Cruise.

www.cnbc.com

The DMV report says "occupants of both vehicles received medical treatment for allegedly minor injuries". It shows one driver and one passenger were injured, presumably the passenger was in the Cruise car.

I think the Prius was legally at fault because it was in the wrong lane and speeding. It's not clear to me why the Cruise attempted a UPL in front of a speeding car, though. Or why it stopped so late in the turn. Seems like it would have been better at that point to proceed out of the intersection. Hopefully the NHTSA will release a report.

 

[Daniel in SD]

NHTSA will investigate the crash last month involving the Cruise AV and a Prius:

U.S. safety regulators to probe crash involving self-driving car from GM-backed Cruise

Federal vehicle safety regulators will investigate a crash last month in which a vehicle struck a self-driving car from General Motors-backed Cruise.

www.cnbc.com

I posted about this here: California Autonomous Vehicle Collision Reports
What doesn't make sense to me is that it sounds like the Prius would have had the right of way if they were turning right but the Cruise vehicle turned in front of them. Perhaps it was predicting that the Prius would slow down to make a right turn? Also, once detecting that the possibility of a collision the Cruise vehicle stopped? It sounds like without that inhuman behavior there wouldn't have been a collision.
Looking forward to seeing the report (in a year or two probably, haha).

 

[diplomat33]

I think the Prius was legally at fault because it was in the wrong lane and speeding. It's not clear to me why the Cruise attempted a UPL in front of a speeding car, though. Or why it stopped so late in the turn. Seems like it would have been better at that point to proceed out of the intersection. Hopefully the NHTSA will release a report.

What doesn't make sense to me is that it sounds like the Prius would have had the right of way if they were turning right but the Cruise vehicle turned in front of them. Perhaps it was predicting that the Prius would slow down to make a right turn? Also, once detecting that the possibility of a collision the Cruise vehicle stopped? It sounds like without that inhuman behavior there wouldn't have been a collision.
Looking forward to seeing the report (in a year or two probably, haha).

Yes, I think the Prius was legally at fault for being in the wrong lane and speeding (2x the speed limit to boot!).

My guess is the Cruise's Prediction and Planning got it wrong because the Prius was in the wrong lane. Since the Prius was in the turning lane, the Cruise assumed that the Prius was going to turn. So the Cruise calculated the trajectory of the Prius based on the assumption it was turning, and estimated that it could safely attempt the UPL. When the Prius did not turn, that completely changed the prediction but it was too late for the Cruise to change course since the Prius was speeding so much.