Why trees have wreaked havoc on Uber's self-driving program


I’m in the software business, and I know from my own experience that you might be expecting to test the hell out of the system in what you expect it to do and where you might guess it will fail, and than that “unknown unknown” shows up to beat you to the ground :slight_smile:

We made some work with researches working in Computer Vision, and separating objects from shadows, especially in the moving picture, is a big deal… or you can get your models trained for the set of objects/conditions you anticipate, then find it reacts incorrectly to otherwise proper inputs, easily classified by the human eye

Here is an article on the similar issue: https://blog.openai.com/adversarial-example-research

Imagine it will get some of the shadows to change a “cyclist” for “trash bin”… sure it’s a stretch example, but explains my point.


ah. So “system” now includes the policy, and the people behind the policy, of where/when/how the cars are used and tested.
And given that “the technology” is not wholly independent of “policy”, and the people behind it, how are you certain that “the technology” shouldn’t be ‘dissed’?


Maybe there’s some ‘supervision’ that let the Computer Vision see that it was a gibbon wearing a panda mask all along! :wink:


In this case yes…It wasn’t a machine that decided to put the technology into the public before it was ready. It wasn’t a machine who decided to only have one human and no way of monitoring him til after the fact.

I admitted that the technology isn’t ready. That doesn’t mean it’s not going to.


I’ve run the gambit of software jobs in my 40+ year career…from software engineer to Senior consultant software engineer to my current position as Director of software development and Senior Principle Architect. Trust me I know software.

Computer vision is only ONE sensor for autonomous vehicles. There are several. As for vision…this technology has made it’s greatest advancements in just the past 5 years. The technology is changing drastically.

Here’s a great article out of MIT

And another one from M Tank…

These are just in the last couple years. Some are saying that computer vision has now exceeded human vision.


then it shouldn’t be out in public where people have been, and can be, harmed


Yes, sure, the fusion of multiple sensors into one model is a way to go, moreover, the “assistant” type of the systems all the major makers introduced moves exactly in the same direction.

Still, it exceeds our abilities in narrow areas, but human brain is far superior in manipulating in wider set of abstractions, as Architect you would know it well :slight_smile:


Are you sure about that? Wikipedia says Waymo had cars on the public roads in Arizona w/o any driver last year, 2017. Here in Silicon Valley it is now legal for driverless cars on the public roads without any drivers at all. This article says Waymo are kicking the safety drivers out in 39 of their driverless cars in a couple of weeks.


So in your estimation it shouldn’t be tested on roadways until it is 100% perfect.
Like the system we have today?
BTW- how do you propose the validation phase of the effort should proceed if it doesn’t involve real world use? At some point, it has to be deployed in the real world because everything else is just a simulation and cannot present all possible scenarios. And, nothing will ever be 100% foolproof. We accept that almost everywhere in technology, what makes a self-driving car different? Take the Lion Air crash for example…is there a cry to remove all planes from use in ferrying people?


Re-read my responses…I said that. I think it’s been pushed ahead of schedule.


Please don’t put words in my mouth – I said nothing about 100% perfect. OTOH, it is entirely possible to discover problems with trees without being on public roads.


too bad you didn’t include that view in your first comment from 3 days ago, or your second and third comments 1 day ago,


This is not the first time Autonomous vehicles have been discussed in this forum. My view on them haven’t changed. I’ve stated more then once over the years that they seem to be rushed to deploy. And re-read my comments on this thread.


Please. You’re the one arguing that the woman got hurt in Arizona and so it shouldn’t be used on the streets. If we remove the ONE instance it becomes perfection.

No, it isn’t. They can test for the potential to detect shadows but until the car is going down the road under all possible conditions, it’s not the same. Anyone who has ever done any type of product testing knows this. That’s kind of why they call it real world testing- because everything else is an approximation…do you honestly think they did not simulate or experience shadowing in their controlled tests?


This is kinda going no where fast . All I know is that the real world testing has to be done and the day will come when I will have to give up driving. And right now I would rather be surrounded by a bunch of self driving vehicles then a lady I know who has had 3 rear end accidents because she will not stop following to close.


Your problem Waterbuff is you’re bashing the technology before it’s ready. It’s still in pre beta test. The technology is still evolving. Systems today are far superior then they were just 6 months ago.

I’m a believer in the technology (or at least the potential of the technology). I firmly believe that any problems they have now can and will be solved.

I have two gripes with the technology though…which has been my position for years.

#1 - Companies are NOT doing enough testing in labs (dedicated roads/cities with no people) before they are put into the public. They said they were ready, and I wasn’t familiar enough with the technology to dispute it.

#2 - The technology is great and is going to drastically makes the roads safer - I’m 100% sure of that. But it’s also going to have a detrimental effect on our economy. It has the potential to drastically reduce jobs in many different sectors…from Police, to insurance to ER doctors to traffic light manufacturers…and the list goes on. One estimate from a professor at MIT’s Sloan School of management says the technology could eliminate anywhere from 50 million to 200 million jobs world wide.


There have always been disruptive technological advances since the dawn of the industrial age. The infamous buggy whip example. Yes, jobs have and will be eliminated. Look how many were created that didn’t exist before this effort to develop autonomous cars. The shift is away from semi-skilled labor so the emphasis needs to be shifted to obtaining these higher level of skills to be employable in the future. This is a natural evolution IMO.


Agreed…but unfortunately the autonomous vehicle could have the largest negative impact on jobs EVER. It has the potential of sending the world into a great depression.


Could happen… but might not.

The same claims were made about trains, electricity, automobiles, airplanes, computers and most of the advances we’ve experienced over the last 200 years. Yet, we’ve continued to grow in global population and prosperity.


It’s only speculation…but it’s not just me or one person. Several world economists have voiced concerned. I will probably never see it. But my kids and their kids will.