FSD video completely fake?

[mongo]

Of course but that’s splitting hairs. Tesla decided to solve rain-sensing differently between AP1 and AP2 cars and Karpathy (amonst others) was tasked to solve this for AP2+ cars. So far they have pretty much failed in my books.

Aye, but sometimes the rabbit must die!

Interestingly for all his prowess Andrej Karpathy hasn’t been able to reach AP1 parity in speed-sign recognition and rain sensing.

AP1 had no rain sensing, so he is beyond parity with the ME system. And current cars could have identical performance, if they had the same sensor. So one could criticize the design decision to no remove it so soon, but one should not bestow upon AP1 features it did not have.

 

[electronblue]

Aye, but sometimes the rabbit must die!

AP1 had no rain sensing, so he is beyond parity with the ME system. And current cars could have identical performance, if they had the same sensor. So one could criticize the design decision to no remove it so soon, but one should not bestow upon AP1 features it did not have.

It was not so much an attempt to bestow the feature on AP1 than to shed light on the limits of even Karpathy’s prowess: speed-sign recognition and reliable rain sensing are still missing.

 

[lunitiks]

Here’s how it works:

The Wiper Thread :-)

 

[firelegend]

The box visualization seems to be a waste CPU cycles.

OK, you are looking for objects by defining outlines, characteristics, and from multiple sensors.

Why wouldn't you just XOR the pixels on the display that triggered your response. This makes the object outline be exactly the opposite shade of the vehicle. So a black car on a black road gives it a white outline. A white car on a concrete road yields a black outline. A white road stripe on a concrete road becomes black outlined stripes around a white center.

What it skips is envelope sizing calculations, and a subroutine for selecting the color, size, and placement of the box and move and size it while still determining borders and features.

I guess it depends on how fast you want your code to run though. The trend today is write bloatware and skip the profiling. Assembler routines and calls? Meh, I gotz ghz, why do I need speed?

I think they do it for themselves. Back when I coded HTML and I was dealing with div's I would place an outline of one pixel in various colors to move things around the page as I needed, sometimes I needed to see how theboxegs were overlapping each other because I didn't use the right attributes etc. Putting the outlines on made it easier for me to see what it was doing. Human eyes need to check how its working, so they can check the accuracy of the IA with their own eyes by reviewing the footage, I would have totally used colors and put outlines on it to "see" what the heck it was doing. Easy enough to turn that stuff on or off.

 

[firelegend]

There is 2017 data, but be aware Google (and others?) moved much of their public road testing to a suburb of Phoenix. It's a fairly nice place for entry level testing. Modern tract homes with good roads. But AV has to work in major cities centers. This is where the initial profits will be. Taxi service in suburbs isn't common. City centers and airports are where taxis get most their money from.

Something isn't right about 2016/2015 though. GM/Cruise was operating both years, first with 5 Leafs, then with 20 Bolts. There should have been more incidents. Cruise started testing on public roads in June 2015, and used different 25 cars prior to Jan 1 2017, and were testing each month ever since. Perhaps incidents were not required to be filed?

But 2017 shows where Google/Waymo reduced participation in California, about April. While there was a single Waymo event in August, it was a situation the AV car was at fault. Debris in the road caused a disengagement as the driver tried to avoid a collision unsuccessfully. Granted there is not always a collision free option with road debris.

In 2016 there were 11 company filings for AV testing in Calif.
In 2017 it jumped to a whopping 43 filings for companies wishing to test in Calif as of Oct 2017.

For more complete info, start here: Testing of Autonomous Vehicles and try to find the right links for various data.

It makes me wonder if companies want to avoid California's very public data releases. This seems unethical if California wants to become a hub of AV testing. By listing AV test performance, it can exaggerate the problems with AV systems. Nearly all collisions have been AV-Not-At-Fault, but they are listed the same. It does not discern between difficult testing and simple testing. Going through the city center of SF is worth the same as a 4 lane deserted road.

But the most significant point of this whole essay is that a huge amount of competitors are working on public road AV testing now, at least in California. Cruise is also operating in Arizona and Michigan and recently started in New York City (which has perhaps the most draconian restrictions, including mandatory police escort).

Who will win this contest? That will not be known for years. It is not rare that the first company to develop a technology ends up losing in the marketplace. However, I think several companies will see various degrees of financial success. Tesla is not going to make ICE cars, and Google/Waymo currently are far behind is automotive production.

My bets on the 3 biggest winners? Google, GM, Tesla. Google will save money on their Google Earth costs, and be able to license their system to other makers, as well as run taxi service. GM has a wide range of vehicles suitable for many tasks up to delivery trucks and route buses and down to subcompacts that could be fitted with Cruise. Tesla will be able to sell their technology to Tesla buyers as well as run Tesla taxis.

My understanding is in some of those cars there are humans in them able to over-ride at any given time.