Thread: Electric cars - Tesla and global

dunno about that my first PC was a 286 xt with CGA monitor and a 20 megabyte HDD

While new computers are not as capable as they ought to be, they do a hell of a lot more than that machine could.

Interesting, 600 usd open-source retrofit autopilot
EON DevKit | comma.ai shop

Originally Posted by EmanuelGoldstein

I think self-driving trucks are stupid. Attack the road-to-track interface problem and use the rails we already have for long distances.


That's possibly not your best example, Bob. Computers of today are no more capable than ENIAC. They are just way faster, and the advance of shit software has made them actually less capable than they were in 1995.

Absolutely. While I have no doubt a self driving truck can handle the highway I am skeptical about how well it would handle some of the trickier situations in congested cities. In the area in which I live there are many businesses that periodically receive deliveries from trucks hauling a 40 foot trailer. Some of the maneuvers they have to make to get into driveways and alleys from narrow crowded streets almost seem impossible.

In one situation I have witnessed more than once, in order to back into an alley the driver must actually drive over the opposite curb, squeezing between a telephone pole and a concrete block building. Then, while backing in he must reposition five or six times, at one point also running some of the trailer wheels over the curb and sidewalk adjacent to the alley. The procedure requires being sure that all vehicles (including bicycles) and pedestrians in both directions have stopped out of the way of the truck's path. All of this without a flagman to control traffic. In order to do this they are breaking several rules (and laws) but police ignore it as they understand the necessity of it.

No doubt a machine could be programmed to duplicate this but the question is SHOULD IT BE? In order for SDVs to operate safely around humans we must program them with certain rules such as remaining within the roadways and not intruding into pedestrian spaces such as sidewalks, except at legitimate spaces where roadways connect. Now, in order to duplicate the feats of the human driver we must teach this dumb machine that on occasions it is OK to violate the safety rules and hope that it will not begin to apply these exceptions in a dangerous manner. A human driver knows the difference whereas the machine is merely performing calculations to accomplish its mission. AI might come up with some wonderful never before thought of solutions to problems but without a human to review the plan before it is implemented we are taking huge and unnecessary risks.



As for our wonderful new computers, EG is correct. They are faster and the software is prettier but in many cases the old DOS programs are actually faster and more efficient.

Originally Posted by MattiJ

Interesting, 600 usd open-source retrofit autopilot
EON DevKit | comma.ai shop

Good God! A cheap way to "hack" the driver assistance features to turn them into a sorta kinda SDV. What could possibly go wrong?

Hint: Think the reason they aren't providing software might be at the demand of their lawyers?

Why do you feel a constant need to dream up various scenarios where a machine might fail? How about if we try to think of some situation where a human might fail? Humans get tired. What if one of them falls asleep while driving? Think anything bad could happen? What if one of them consumes alcohol before driving? And on and on and on. Apparently you are strangely OK with all this risks.


Sent from my iPhone using Tapatalk

Hell, what about the REALLY scary scenario? EVERYTHING GOES GREAT! Nothing to worry about! Working as intended. Very few people seem to even want to look at that.

Originally Posted by Scottl

.....
As for our wonderful new computers, EG is correct. They are faster and the software is prettier but in many cases the old DOS programs are actually faster and more efficient.

I get that you want better. We all do.
Assume you mean PC or MS-DOS which did not exist 40 years ago.
My CDC-6500 and IBM-370 could not even spell check my card deck let alone auto-correct it for me.
It just died, spit you out and you waited in the queue for up to 24 hours after submitting another deck for one typo which it would not point out.
In those days machining was paper tape and you had to cut and splice hoping it would feed part after part.

I think it has changed a lot over my lifetime. As far as smarter look at what the engines can now do in seconds or hours in the AI field.
If you coded the same of a SDV on a 40 year old system 40 years ago it would just about now be solving it.
Even if I could fit a 370 or a CDC in my car that reaction time is a tad slow.

Yes, all this free cpu time and speed has led to bad programing for the masses but real time programming and desktop user applications are as different as C++ and Gcode.
Forty years back what was your CAD/CAM and what did it run on? Expensive, very custom terminals?
Lets do thirty years back as the personal computer and DOS where in the running.
Bob

Originally Posted by adh2000

Why do you feel a constant need to dream up various scenarios where a machine might fail? How about if we try to think of some situation where a human might fail? Humans get tired. What if one of them falls asleep while driving? Think anything bad could happen? What if one of them consumes alcohol before driving? And on and on and on. Apparently you are strangely OK with all this risks.


Sent from my iPhone using Tapatalk

It isn't just me questioning it. A lot of people who are actually working on the technology say true autonomy in all situations (Level 5) is a long way away. Too many with dollar signs in their eyes are telling us "Don't worry, be happy" and pointing to success in controlled situations while ignoring all those REAL WORLD situations such as what I have described. Do you really understand what AI is (and more importantly isn't) and the limits of machine vision in the real world? outside of limited access highways and select city campuses there is a lot of unpredictability faced by drivers. Can the machine see the subtle hints that a pedestrian is about to abruptly change direction and dart into the street? It happens here all the time. Just as importantly, can it respond to unique situations quickly enough to not be a bottleneck to traffic? Periodically I have to travel into the Boston City Hall/Government Center area and then walk several blocks to my destination. Pedestrians are supposed to wait for crossing lights and cars aren't supposed to proceed when people are in the crosswalk BUT - It happens all the time, every day. It looks like an elaborate dance, with vehicles passing through groups of pedestrians, but it works because all but out of towners understand the game. It's all illegal as hell but it's how things keep moving. Can your precious SDV be programmed to understand the dance? Can you teach it about the special hazards of dogs and kids on the sidewalk and how often they suddenly dart into the street?

Full Page Reload

Warning: Driverless Cars Are Farther Than They Appear

Note what this optimistic prediction below mentions about the need for "AV-specific highway lanes" and "fully automated vehicles operating in geofenced areas along dedicated routes". That is NOT true level 5 autonomy, which means operating in our world without special dedicated (and expensive) infrastructure. It's bad enough that we have lanes dedicated to High Occupancy Vehicles but next they are demanding we give up yet another lane for SDVs (and no mention of how both they and we are to share on and off ramps.

Are Autonomous Vehicles Farther Away Than We Think?

Originally Posted by Scottl

.. Can you teach it about the special hazards of dogs and kids on the sidewalk and how often they suddenly dart into the street?

That is NOT true level 5 autonomy, which means operating in our world without special dedicated (and expensive) infrastructure.

On line one you do not really teach it but this is a big thing to get in the system. You do not want to make mistakes here. Certainly this should be in there which is why we build test tracks that do just that.
On line two I'm not sure level five is ever possible against a profession driver. But the world is not full of pro race car drivers.
(My wife just recently proved this and I was so puking upset with her but I have to realize that this is her best under the conditions she saw)

The hope is to reduce the damage and fatalities that occur everyday.
If your SDV has five times less crashes but it still kills some people is it better?
Bob

Now anyone can have any opinion, but when they expose it publicly it should be defensible, by them, according to common metrics.

Tesla has about 500k cars sold, on the road, and maybe 150k-200k with autopilot enabled == paid for.
About 30k+ cars/month are delivered every month.
About 20-30%-40% of them choose to buy autopilot, depending.

In 12/2019, we have no reports of great numbers of monthly accidents, and/or deaths, due to Tesla autopilot.
According to every actual data point available, the Tesla autopilot has less accidents and deaths than human drivers, by mile travelled.
By about 50%, or twice-as-safe, at the very least.
When comparing tesla-without and tesla-with ap, so apples to apples.

No doubt several very public fatalities have occurred with Teslas and/or ap, maybe about 12 globally, past 5 years +/-.
These were tragic.
But less than similar miles driven on similar cars, globally.
And the nr of miles driven, per fatality, and the number of cars, per mile, is increasing exponentially every 14-16 months.