Slashdot Mirror
People Are Losing Faith In Self-Driving Cars Following Recent Fatal Crashes (mashable.com)
oldgraybeard shares a report from Mashable: A new survey (PDF) released Tuesday by the American Automobile Association found that 73 percent of American drivers are scared to ride in an autonomous vehicle. That figure is up 10 percent from the end of last year. The millennial demographic has been the most affected, according to the survey of more than 1,000 drivers. From that age group, 64 percent said they're too afraid to ride in an autonomous vehicle, up from 49 percent -- making it the biggest increase of any age group surveyed.
"There are news articles about the trust levels in self-driving cars going down," writes oldgraybeard. "As a technical person, I have always thought the road to driverless cars would be longer than most were talking about. What are your thoughts? As an individual with eye problems, I do like the idea. But technology is not as good as some think."
The Mashable article also references a separate study from market research company Morning Consult "showing increased fear about self-driving vehicles following the deadly March crashes in the Bay Area and Arizona." Another survey from car shopping site CarGurus set to be released Wednesday found that car owners aren't quite ready to trade their conventional vehicles for self-driving ones. "Some 84 percent of the 1,873 U.S. car owners surveyed in April said they were unlikely to own a self-driving car in the next five years," reports Mashable. "79 percent of respondents said they were not excited about the new technology."
The Mashable article also references a separate study from market research company Morning Consult "showing increased fear about self-driving vehicles following the deadly March crashes in the Bay Area and Arizona." Another survey from car shopping site CarGurus set to be released Wednesday found that car owners aren't quite ready to trade their conventional vehicles for self-driving ones. "Some 84 percent of the 1,873 U.S. car owners surveyed in April said they were unlikely to own a self-driving car in the next five years," reports Mashable. "79 percent of respondents said they were not excited about the new technology."
The link is to a local file, not net-accessible....
I do industrial automation for a living, since about 2000. There's a certain class of automation problem where getting to a 90% solution is easy, getting to 95% takes a lot of work, and getting to 97% is extremely hard. That is, 90% of the parts coming down the assembly line are easy to categorize correctly, the next 5% you can do with a lot of effort, and so on. Unfortunately that last 2 or 3% are damn near impossible due to problems with how good our sensors are, or how good our algorithms are, or how good our mechanical sorting solutions are.
These problems are notorious for causing run-on projects that slurp up money but never end. That's because your initial effort appears to produce amazing results - 90% with almost no effort. How hard can the remaining 10% be? My first encounter with one of these problems was a barcode-reading system at an industrial facility reading barcoded tags with a camera instead of a barcode reader. The problem was that the barcodes were becoming more worn and faded over time, and management believed that if we used a camera instead of a barcode reader we'd be able to enhance the image, etc., and get a good read because clearly a human looking at the picture can clearly see the bars and the human-readable text below it. This project went on for months, and then years, always creeping closer to 100%, but never making that leap to 100%, having thrown several different engineers at the problem and bringing in outside machine vision specialists.
In most cases these problems come from over-estimating the capability of your sensors. A sensor with a little dirt on it suddenly gives the wrong result, or temperature fluctuations mess up the calibration, or the dreaded, "sensor seems to be giving valid values, but they're just wrong for no reason." Even if your sensor values are reliable, in many cases you'll end up with a measurement that doesn't fall clearly into the known-A or known-B range.
That's where "AI" is supposed to save us, but my limited experience with AI shows it falls into the same class of engineering problem: you can quickly build an AI that correctly categorizes 90% of your input correctly, and then with effort you can improve it and improve it some more, but you'll never reach that always-correct answer.
This is where engineering projects fail, because you can always find a manager or an optimistic engineer who can hand-wave away the ambiguity and say, "humans aren't perfect either" and "we can just keep making the AI better and better." That's convenient when you don't put a physical number on it. How good can you make the AI with the available sensors? We know the sensors are in some ways better than human perception, but in other ways they're worse. In what quantitative ways are they worse, and how are you compensating for that?
If I were going to tackle some problem like this, I'd start with a standardized sensor suite and data format. You can't have everyone developing AI based on proprietary sensor data because it's too opaque. You also need to standardize the system output format (accelerator percent, braking percent, steering value, etc.) Plus you need to standardize the parameters of the vehicle. Once you've got that you need to start collecting and publishing this data in this standard format - hundreds of thousands or millions of test case scenarios available for every researcher to use, and in each case you need to have an expert specify what the correct set of outputs should be (or correct range at least) for each scenario. Then you can develop your AI or algorithms and you can then run these through a test suite so your AI has to pass all of these scenarios before it can be certified. As we have crashes then we add to the list of scenarios, and if you make changes to the AI, it has to pass that new scenario and still pass all the old ones.
I get the sense this is what the companies doing research are trying to do, but how do we validate their product? If their databases are proprietary, and their sensor format and data isn't in a standard format, and we can't run the tests ourselves, then how can we trust their systems? Of course we can't.
"I have never let my schooling interfere with my education." - Mark Twain
Are these reporters pointing out that 17 gasoline cars burst into flames every hour in the USA? That non-Tesla cars are responsible for 6% of all fire-related deaths?
Nope? That's what I imagined.
https://www.nfpa.org/Public-Ed...
No sig today...
The issue is the various tech. Camera only models will have drastically reduced abilities compared to lidar,radar,camera models.
Self driving cars can't go the cheap route like Tesla autopilot. You need the $125,000 package of equipment.
i thought once I was found, but it was only a dream.
As someone who drives a Tesla with âoeautopilotâ features, I believe full self-drive is a long way off. For a start, from time to time, the Tesla does veer over the middle divider line, or worse, over the line at the edge of the road. On a long drive on a road with some curves, I expect this to happen once or twice. So even that stuff is not reliable. Heck if youâ(TM)re coming down a hill, the sensors donâ(TM)t even see the car in front of you sometimes because of the angles....going around a corner where there are âoesuddenlyâ stopped cars because of a traffic light is another issue....car only notices at the last moment! But the bigger is issue is anticipation. If Iâ(TM)m driving on a street and there are kids playing with a football on the sidewalk, I know it makes sense to slow down, move a little further out into the road, just in case one of the kids runs out to get the ball. Or I see a truck stopped and I know thereâ(TM)s a possibility the driver might open the door, etc. All these self-drive systems are reactive and I donâ(TM)t think thatâ(TM)s good enough for safe driving, even compared to people.
Humans drive 3.22 trillion miles a year in the US, in 2010 there were 5.5 million crashes This includes all claimed fender-benders, in all driving conditions in the US. This means that they are out there driving over 585,000 miles successfully per crash. I think Waymo has maybe achieved 5700 miles per 'interaction' which is the measure the industry has chosen to indicate a 'crash'.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
This might help in the future once standards are defined and implemented. Currently, at least on the waymo front, I know that Google has been using captchas to train AI to recognize signs and vehicles. Everyone has seen these captchas.