Why Full Self Driving is needed

Sure.
However, stopping people from speeding would save 100x as many lives as stopping people from rolling through a stop sign WHEN NO PEOPLE OR CARS ARE DETECTED NEARBY.

Mike

 

And why weren't these issues addressed and fixed during internal testing? You beta something that is not going to feature change before general release. But this is out on our highways in tens of thousands of cars. I get the "the more data the better we can change and the next release will be better for it" philosophy, I just question if it is appropriate when lives are at stake.

 

What's disheartening is that they are using beta testing as a substitute for professional unit AND integration testing.

But even worse is that the programming team got the OK to deliberately break the laws of all 50 states. As I watched the videos that were posted, I saw that the car was changing lanes in the middle of intersections (illegal) and driving for for an extended period along the curb in the parking spaces. Someone, somewhere either screwed up their testing or signed off on those behaviors.

I would not ride in a Tesla. The software is too slipshod, and it appears that their management team is signing off on shoddy work. As a systems analyst + programmer, I would have been fired if I put a product like that in the hands of the public.

 

Well, that is your opinion and you are entitled to it.
But we are talking about features that are completely optional to use and even ones that the driver had to explicitly qualify for, ask for and agree to use cautiously.
Now certainly there will be people to choose to drive recklessly anyway.
But then people of all makes and models choose to drive recklessly and you are not asking them to stay off the roads, are you.
We have the technology to monitor them and take them off the roads, right?

You seem overly worried about Tesla cars driving around town at 30 mph.
Speed kills far more people than Teslas. (Maybe DUI is more?) Are you advocating for zero tolerance? One speeding ticket > 10mph over the limit and you lose your license and impound your car.
Or what about disallowing fast cars all together. Why do we allow a BMW to go over 100 mph?
Why do we allow GM to make the Corvette go 184 mph?
Why are you not complaining and having them recall it and ask how did management sign off on this?

None of this is meant to allow Tesla to do stupid of unsafe stuff...but it ranks pretty low compared to this list
The Corvette is only #6 on this list
https://www.hotcars.com/most-dangerous-cars-ever/

Mike

 

If you don't think that deliberately coding the FSD logic to perform illegal maneuvers is stupid or unsafe, you have another think coming.

It's amazing that you are aware of all of my opinions and complaints. Sadly, your omniscience is not as complete as you think it is.

If I had the authority, I would probably impose reasonable speed governors, tied to local speed limits, and zero tolerance for people who manage to exceed the limits. DUIs would be one strike and you're out. After all, every drunk driver knows in advance that they will be driving a multi ton machine to get home and chose to drink anyway.

You are far off center if you think that the more dangerous features are only active when you chose them to be. If they can't program the car to avoid parked emergency vehicles, why would you think that they could prevent the code from accidentally triggering the self driving routines?

 

If I don't trust Tesla with anything else, why would I expect them to be truthful about the crash data that they collected?

As a systems programmer, Bob, you should be seeing a big red flag when they threw the lane keeping software into the same bucket of data as the data for the FSD driven cars. Another big flag goes up when they then compared that against the NHTSA's stats concerning accidents for ALL cars, regardless of age, condition or weather.

It's pretty clear that Tesla is playing fast and loose with the stats in the hope that they can make FSD look good.

 

They ought to be doing that too. It isn't like they can't address more than one unsafe and illegal practice. But that is for a separate thread, so go ahead and start one.

I notice that you chose to say "detected", not "present". Given Tesla's driver-assist crash history of not detecting barricades, truck trailers, and emergency vehicles stopped along the road, this gives me no comfort. We need to be aware of 'false negatives' and other errors and failures, and use multiple safety layers so that a single failure in one layer doesn't lead to disaster. Applying multiple safety layers is a big part of how commercial aviation achieved its vastly improved safety record. Failure to obey multiple overlapping safety rules is how a movie producer/actor recently failed to "detect" a live round in a prop and ended up killing his camerawoman. Had he followed all, or even any, of the additional standard safety layers, she would have been uninjured.

BTW, the news articles say "no “relevant” moving cars, pedestrians or bicyclists can be detected", which is a lesser standard than the "no[ne] detected" that you asserted. This adds to the concern. Are all detected items that are dismissed as irrelevant, done so correctly?

5 mph, or 5.6 from the news articles and FW release notes, is a jogging pace, almost double the standard walking pace of 3 mph.

Does Tesla properly address sampling bias? E.g. are there any roads or conditions that are unsuitable for AP, and thus require the driver to handle unassisted? (gravel and other roads without any or adequate lane markings, foul weather, etc.)

A proper comparison between AP and unassisted modes would include miles driven only on roads and in conditions suitable for AP. And probably weighted for differing amounts in differing conditions. Unassisted miles and crashes on unsuitable roads must be excluded, otherwise the comparison is not apples-to-apples. Cherry picking. Because driver assists started with the very easiest conditions, and gradually expanded towards tougher conditions, I would not be the least bit surprised if still-unassistable conditions naturally have higher crash rates.

 

Speaking of unsafe, #11 on this list (from the same click-bait site) is relevant to this thread:
https://www.hotcars.com/ranking-15-of-the-most-dangerous-driving-habits/

(FWIW, I have not problem riding in a Tesla. Just how the driver makes use of the various assist features is an entirely different matter.)

 

I also think is was dumb for them to do this!
But more from a PR point of view than a safety point of view
I just don't think it is nearly as dangerous as driving fast, DUI, driving sleepy, etc etc.
Human drivers regularly roll through stops and accident rates (certainly deaths) is very low by comparison

Driving with FSD requires that you keep your hands on the wheel and eyes on the road.
Drivers had to drive very safely and get accepted into the program etc.
The cabin camera does watch and make sure you are paying attention
The car driving down the road is not an autonomous car...it is an autonomous car with a human immediately ready to correct any errors.
Drivers not doing this are just plain reckless and/or looking for clickbait on Youtube.

Generally the FSD is too timid and conservative and to keep from annoying other drivers ...I have to frequently tap the accelerator (to tell it is OK to go) so I don't block traffic. It is probably going to take a year or two to just get to be usable as Level 2. (which requires hands on wheel and eyes on the road), just not "beta"

Mike

 

https://www.insideevsforum.com/community/index.php?threads/autopilot-accident.7438/

 

We really shouldn't be calling them autonomous at all then. It leads to confusion, and users overestimating what their car can do. These are just driver aids. The help ease the burden of driving, but the operator has to still be paying full attention as when controlling everything.

 

I don't think you can say that they "fixed that problem" since what they did was to throw out the old navigation engine (using radar and visible light) that was used a couple years ago and built new software based only on a neural net's analysis of captured video.

One of the things I dislike about machine learning is that it is only as good as the rules supplied and the accuracy of the training cases that it tried to learn from. If no one bothers to include people in wheel chairs in the sample dataset for "pedestrians" there is a good chance that a person will be run over at some point. Best case would be that the handicapped would be regarded as an immovable object like a flower pot.