There were many published reports of Waymos encroaching on crossing guards and children in crosswalks in San Francisco. More evidence of Waymo prioritizing performance over safety. I wonder if the recent acknowledgement of Waymo's failure to operate safely around school buses is related to nascent operations in the less dense cities of Austin and Atlanta where school buses are more common.
To Henry English, I suggest you read Waymo's Terms of Service. Once you see how they shift the burden of liability onto their customers and deny customer access to the legal system in favor of forced arbitration, you will find another very good reason to take a taxi instead, as I did in Phoenix recently. Interestingly, on my way to the Phoenix airport an unoccupied Waymo attempted to cut off the taxi, pulling away from the curb and honking at the taxi to get out of its way before jamming on its own brakes. Perhaps word of my actually reading the Waymo terms of service and refusing the Waymo ride has gotten around in the Matrix.
I think you did a pretty good job with a couple exceptions
You did not mention a larger systemic issue with these only being a symptom. Not just based on the issues Phil discusses, but others. Phil still does not get or say the industry uses a development approach that is not viable from a safety, time or cost POV. More to follow.
These are systemic not edge cases or one offs-Humans have no issues here - They are lying about being anything close to "fully autonomous"
· Refused judge’s request for # of vehicles in SF outage
· Freezing in San Francisco power outage
· Driving down train tracks – Riders had to bail out
· Passing buses picking up/dropping off kids-The fix failed
· Driving down the wrong side of the road into traffic
· Hitting a pole and blaming mapping
Hopefully you get this now?
No one in the industry is using a viable development approach. No one can spend the lives, money or time to get near L4. (Which is why we see zero meaningful test data) Each of these need to be resolved to reach L4:
-No General Learning/inference. Results in an untenable pattern recognition effort. Also, "pixel" or micro level classification is often wrong when only a few pixels are in contention
- Use of needless human Guinea pigs. It is virtually impossible to regain situational awareness in time critical disengagement scenarios like crashes.
- Relying on the real-world vs simulation for most development
- Use of inadequate gaming sim tech and sensor modelling vs what aerospace uses
- The 360 deg sensor design needs to be=Cameras, LiDARs and imaging radars. Each providing classification, localization and tracking to assure primary and secondary modal operation through the ODD, especially as conditions degrade. (And sound localization)
Regarding data
No one has seen ANY meaningful data by design from anyone. Data that ensures the engineering approach weaknesses of pattern recognition AI/ML and pixel level DL are mitigated. The purpose here being to hide what is really going on and mislead. The most important data being how often the human “safety driver” intervenes and prevents crashes. Beyond that is the hundreds of millions of single and combined learned object, scenario and environment variations, especially for crashes and bad weather and lighting. Focusing on the benign is dangerously misleading. Beyond that, Rand said it would take 100B miles to prove a system is as good as a human. Developing the system with non-GL AI/DL, which is pattern based down to the pixel level for classification, would require so much work it would be impossible to do. And that assumes you are god and can make the exact same thing happen over and over. Where simulation is leveraged, few model sensors. When they do they are low fidelity and will create more false confidence than not because of poor simulation technology.
Phil, thank you for going down in deep on this topic, and making practical suggestions as to what Waymo can still do to redeem themselves. An important topic that we should all take a note of.
Phil, enjoyed your post and I agree Waymo needs to fix this fixed.
You may call this whataboutism, but I’d like to understand better how you think about these safety issues.
Humans have a major problem with speeding. Just about everyone does it, on every street, everyday. Speeding is clearly linked to an increased chance of accidents happening.
Should anything be done about this? The existing system of tickets and license points has failed to curb this behavior and people are dying every day as a result. Are any reforms needed?
Thanks, not trying to troll, trying to understand what’s a problem and what’s acceptable.
How about all the people Robo taxis will put out of work? I for one Will take human driven taxis on that account. Alternatively, how about Robo taxis being required to have human “Driver supervisors”? I for one will take human driven taxis on that account. Alternatively, how about Robo taxis being required to have human “Driver supervisors“? You would have the best of both worlds: People having jobs and improved safety. Of course, that would defeat the purpose: the whole idea is not having to pay salaries and benefits.
Thanks for the question. Asking 'how can humans be better' is a very important question so long as it is not used to justify robotaxis making avoidable mistakes.
Europe has done a much better job at road safety than the US, without robotaxis.
Here is a starting point, and for more look into Vision Zero.
And there’s one solution that can be implemented easily. Waymo can bout and install IoT devices for school buses.
Indeed, if they can't detect them any other way they might offer to install (at Waymo's expense) transponders on the buses.
Apparently many school districts already have tracking apps for parents as well. Apparently Austin ISD has had such an app since 2016: https://www.kxan.com/news/new-app-allows-austin-isd-parents-to-track-their-childs-bus/
Thanks, Phil.
There were many published reports of Waymos encroaching on crossing guards and children in crosswalks in San Francisco. More evidence of Waymo prioritizing performance over safety. I wonder if the recent acknowledgement of Waymo's failure to operate safely around school buses is related to nascent operations in the less dense cities of Austin and Atlanta where school buses are more common.
To Henry English, I suggest you read Waymo's Terms of Service. Once you see how they shift the burden of liability onto their customers and deny customer access to the legal system in favor of forced arbitration, you will find another very good reason to take a taxi instead, as I did in Phoenix recently. Interestingly, on my way to the Phoenix airport an unoccupied Waymo attempted to cut off the taxi, pulling away from the curb and honking at the taxi to get out of its way before jamming on its own brakes. Perhaps word of my actually reading the Waymo terms of service and refusing the Waymo ride has gotten around in the Matrix.
As I have said so often in the past... just added to my syllabus. Right next to William Widen's video on "Law as a Design Consideration in AVs".
For my Detroit-based auto engineering managers. Maybe it will have an impact? 🤷♂️
Phil
I think you did a pretty good job with a couple exceptions
You did not mention a larger systemic issue with these only being a symptom. Not just based on the issues Phil discusses, but others. Phil still does not get or say the industry uses a development approach that is not viable from a safety, time or cost POV. More to follow.
These are systemic not edge cases or one offs-Humans have no issues here - They are lying about being anything close to "fully autonomous"
· Refused judge’s request for # of vehicles in SF outage
· Freezing in San Francisco power outage
· Driving down train tracks – Riders had to bail out
· Passing buses picking up/dropping off kids-The fix failed
· Driving down the wrong side of the road into traffic
· Hitting a pole and blaming mapping
Hopefully you get this now?
No one in the industry is using a viable development approach. No one can spend the lives, money or time to get near L4. (Which is why we see zero meaningful test data) Each of these need to be resolved to reach L4:
-No General Learning/inference. Results in an untenable pattern recognition effort. Also, "pixel" or micro level classification is often wrong when only a few pixels are in contention
- Use of needless human Guinea pigs. It is virtually impossible to regain situational awareness in time critical disengagement scenarios like crashes.
- Relying on the real-world vs simulation for most development
- Use of inadequate gaming sim tech and sensor modelling vs what aerospace uses
- The 360 deg sensor design needs to be=Cameras, LiDARs and imaging radars. Each providing classification, localization and tracking to assure primary and secondary modal operation through the ODD, especially as conditions degrade. (And sound localization)
Regarding data
No one has seen ANY meaningful data by design from anyone. Data that ensures the engineering approach weaknesses of pattern recognition AI/ML and pixel level DL are mitigated. The purpose here being to hide what is really going on and mislead. The most important data being how often the human “safety driver” intervenes and prevents crashes. Beyond that is the hundreds of millions of single and combined learned object, scenario and environment variations, especially for crashes and bad weather and lighting. Focusing on the benign is dangerously misleading. Beyond that, Rand said it would take 100B miles to prove a system is as good as a human. Developing the system with non-GL AI/DL, which is pattern based down to the pixel level for classification, would require so much work it would be impossible to do. And that assumes you are god and can make the exact same thing happen over and over. Where simulation is leveraged, few model sensors. When they do they are low fidelity and will create more false confidence than not because of poor simulation technology.
Phil, thank you for going down in deep on this topic, and making practical suggestions as to what Waymo can still do to redeem themselves. An important topic that we should all take a note of.
Phil, enjoyed your post and I agree Waymo needs to fix this fixed.
You may call this whataboutism, but I’d like to understand better how you think about these safety issues.
Humans have a major problem with speeding. Just about everyone does it, on every street, everyday. Speeding is clearly linked to an increased chance of accidents happening.
Should anything be done about this? The existing system of tickets and license points has failed to curb this behavior and people are dying every day as a result. Are any reforms needed?
Thanks, not trying to troll, trying to understand what’s a problem and what’s acceptable.
How about all the people Robo taxis will put out of work? I for one Will take human driven taxis on that account. Alternatively, how about Robo taxis being required to have human “Driver supervisors”? I for one will take human driven taxis on that account. Alternatively, how about Robo taxis being required to have human “Driver supervisors“? You would have the best of both worlds: People having jobs and improved safety. Of course, that would defeat the purpose: the whole idea is not having to pay salaries and benefits.
Thanks for the question. Asking 'how can humans be better' is a very important question so long as it is not used to justify robotaxis making avoidable mistakes.
Europe has done a much better job at road safety than the US, without robotaxis.
Here is a starting point, and for more look into Vision Zero.
Seven minute video on people's driving and the fallacy that robotaxis are the one true answer to that: https://youtu.be/pYb4X5aJhgU?si=7vchWvBF4ZUpo8hu