Forums

Full Self Driving Vehicles are Here Already *UPDATED*

Full Self Driving Vehicles are Here Already *UPDATED*

Tesla employees are now Testing FSD

https://www.bloomberg.com/amp/news/articles/2018-09-28/tesla-enlists-emp...

Elon Musk has asked for Tesla Inc. employees to test what the company has billed as full self-driving capability and is dangling $13,000 in savings to entice them to help.

Musk wrote in an email obtained by Bloomberg News that Tesla needed about 100 more employees to join an internal testing program linked to rolling out the full self-driving capability. Any worker who buys a Tesla and agrees to share 300 to 400 hours of driving feedback with the company’s Autopilot team by the end of next year won’t have to pay for full self-driving -- an $8,000 saving -- or for a premium interior, which normally costs $5,000, Musk wrote.

+++++

Since March 2018, the Neuhausen Swiss transit authority has been operating a Navya 11-passenger autonomous bus (nicknamed “Trapizio”) on a regular fixed route, integrated into its public transit bus system.

Instead of a driver, the vehicle hosts a “steward” who welcomes passengers and can control the vehicle if necessary as it makes a 10-minute loop through town. After 8,000 rides on Trapizio, the city is already looking to add a second route.

https://www.axios.com/swiss-town-making-way-for-avs-in-public-transit-77...

I'm tired of those who claim without evidence that full self-driving vehicles are 5-10 years away. They are wrong. It doesn't matter how or why, but these vehicles are here operating safely.

I'm less concerned with SAE levels or geographical or conditional operation.

TESLA NETWORK:

Tesla Network will provide point-to-point operation for certain geography first. Not all over the earth at the same time. States like California, Nevada, Arizona, and foreign locales like Switzerland are already adopting these safety-first operations. It is only a matter of time before Tesla activates their Uber/Lyft/Waymo network of 100% electric vehicles with eventual rollout to areas that adopt laws regulating FSD vehicles. With a network of super-fast chargers. Already powered by sunlight and stationary batteries at some locations.

Nobody is saying other companies won't eventually get there, but Tesla has already sold ~200,000 FSD-equipped, data-gathering robots. And 8,000 more every week. Waymo will "eventually" buy 20,000 EVs from Jag/Minivans? How much is the extra ugly hardware/Lidar? That's 2 weeks of Tesla production by the end of 2018. Tesla has a walled garden the likes of which Apple WISHES it could get.

#1 selling Full-Size Sedan

#1 selling Luxury SUV

#1 highest grossing Sedan.

https://1reddrop.com/2018/08/24/tesla-model-3-highest-grosser-u-s-beats-...

Most people I know admire the underdog that fights GM/FORD/VW/Mercedes, Oil/Gas/Coal/Utilities, Texas/Car Sellers/Presidential Candidates, China/Russia, Defense Contractors and Wall Street and Tesla is in for one hell of a fight.

EVRider | 26/08/2018

As far as I know, all of the existing autonomous driving systems operate within certain constraints that will not be there for autonomous passenger cars. FSD has to work anywhere and everywhere.

SamO | 26/08/2018

@EVRider,

That's simply false. FSD could work in Los Angeles County or within the city limits of Phoenix. It would still be full self driving, with geographical/legal limitations.

carlk | 26/08/2018

Like EVRider said all those are geofenced systems on pre-mapped roads only. The only exception is what Tesla is developing and only because it does not use LIDAR. Tesla is able to put FSD hardware in hundreds of thousand cars to collect billions of machine learning miles a year. Others can only get a small fraction of that in a few cities at best with their test fleets. Some like Waymo also do "simulated" run although that could never match data collected from the real world where cars are going to be.

.

carlk | 26/08/2018

Of course the smaller area the FSD car run the easier it is to make it to work. That's why all those autonomous vehicles today are for things like airport or campus transportation. Only Tesla will provide a car you can buy and tell it to go anywhere you want it to go without your intervention. Others, Waymo included, don't even have a path to go there.

Remnant | 27/08/2018

@carlk (August 26, 2018)

<< Like EVRider said all those are geofenced systems on pre-mapped roads only. ... Tesla is able to put FSD hardware in hundreds of thousand cars to collect billions of machine learning miles a year. >>

Indeed, it's real time re-mapping and traffic updating that FSD must have, along with the optimal algorithms, in order to meet the drivers' expectations AND needs. Whence, vast additional connectivity AND processing capabilities MUST be available in the FSD HW+SW to function properly.

Go Tesla ... !!!!!!!!!!!

bp | 27/08/2018

A major challenge for FSD is handling all of the unusual situations that can occur when driving at posted speed limits on roads where something unexpected can happen, and the FSD software must be able to react quickly enough and in a manner that is at least as safe as an experienced human driver.

It's much more than maintaining the speed limit & spacing behind the vehicle ahead, detecting traffic signs/signals/speed limits, lane keeping, and following a navigation route.

All current FSD implementations have significant limitations to simplify the environment enough so the FSD software can navigate safely - such as limiting the geographic area, running in areas with limited traffics/objects, and/or willing to slow down or stop when encountering something the FSD software can't handle.

While access to cloud data and processing can help in getting up-to-date conditions and optimize routing, access to network resources can't be guaranteed 100% of the time, nor can it be relied on to quickly react to driving situations - which is why Tesla's planning to upgrade the AP processor next year.

For Tesla FSD customers, what we'll likely see is continuing improvements in "driver assist" features, with the software operating for longer periods without requiring human overrides and in following the navigation routes for more of the route going from point A to point B, navigating through turns, highway entry/exit, lane changes, etc. - likely working better (at least initially) on limited access highways.

And even if we don't get to a point where we can run without "hands on wheel" for a while, that should provide enough value to justify FSD activation.

carlk | 27/08/2018

@Remnant

Yes it's getting clearer everyday Tesla's approach is the ONLY way that will work for general FSD. We can also see now Tesla actually has planned all this from the beginning years ago when it put the hardware in every car whether you pay for it or not. Once again it proves the power of first principle of engineering. LIDAR sounds good on paper but the way it can only be implemented on limited test cars will never get us there. Elon said LIDAR is like on crutches. Waymo or all its imitators will have to stay with that limitation until when they decide to go with the Tesla approach. But then they will have to start over again. Everything they have obtained with LIDAR will no longer be useful. People only look at things by analogy will never see that Tesla indeed has years of lead in this field.

mos6507 | 27/08/2018

[Only Tesla will provide a car you can buy and tell it to go anywhere you want it to go without your intervention.]

will = future-tense.

We're not there yet. You're just slinging a prediction. Nothing more.

carlk | 27/08/2018

Also other companies seem to be so hung up on "eyes" but Tesla knew it's the "brain" that matters. That showed in the recent announcement that it will produce its own more powerful AI chip, a project that was also started years ago. Sensor system on Tesla already has better capability than the human eye. There is no reason to use LIDAR and having to restrict how fast the brain could learn. Making the "brain" better and to make it to learn faster makes much more sense. Human don't born with the ability to drive even after they have developed the perfect vision.

I still remember arguments a couple years ago whether Tesla dumped Mobileye or Mobileye dumped Tesla. Again it's clear now that Tesla dumped Mobileye. Tesla was thinking way ahead of what Mobileye could ever do.

carlk | 27/08/2018

@mos

You understand the first principle of thinking?

Like I said sensor suit in Tesla cars is already better than the human eye. NVidia PX2 has 20 billion transistors versus ~100 billion neurons in human brains with which only part of them are used to process vision and driving related functions. The new Tesla chip is said to be able to process images ten times faster. And even that is not a limitation as better chips can always be made in the future. All it takes is enough time to learn as we all have too learn how to drive. There is no reason, again if you can follow the first principle, that this approach will not work.

SamO | 27/08/2018

No @mos doesn't understand thinking of any sort. He's been spectacularly wrong about every single position he's take regarding Tesla.

TeslaTap.com | 27/08/2018

I agree LIDAR is a crutch when you have unlimited budget, but LIDAR prices are coming down. Still requires massive additional compute and you still need vision. I think it's too early to conclude one approach or another is best. For today and likely then next 2 years, vision only seems to be the best practical approach to make a car mere mortals can afford. Waymo can afford adding $50-100K to each car for testing. Less clear when prices will be in the $1-2K range. A few $100-200K cars have been announced (but not delivered) with LIDAR, with a lot of hopes the prices will be reasonable and they will actually work in the rough automotive environment.

I agree with others here, vision can do it without LIDAR, with a huge cost advantage even if LIDAR gets cheaper. There is talk about $200 solid state LIDAR units (4 required - so $800), but these are not available yet, and still require some kind of mechanical wipers to keep clean adding to the cost/complexity. General thought is to install 2 of them in the headlights and have headlight wipers.

I'm sure eventually LIDAR can be made to work at a somewhat economical price - but it will alway be significantly more expensive than vision only system, and it's not clear it provides any improvements to FSD operations.

georgehawley.fl.us | 27/08/2018

Sorry to disagree, @SamO but I feel general purpose FSD is still more than a year or two away. Summon won't even work well enough to back my Model 3 out of my garage because the opening is too narrow.

I live in a gated community. The streets have no lane markings. Entering the property requires recognizing a laser scanner and driving close enough to it for it to work. It requires the car to be stopped short of a horizontal gatepost that the car cannot detect at this point. The car has to remember which garage door to use and to park in the left 2/3 of the space on the right side of the garage. These are all solvable but illustrative of the many complexities confronting the Tesla SW engineers. As I drive around South Florida I am now sensitive to the conditions that present obstacles for successful FSD. There are many and some of them are daunting. There is a long stretch of I4 in Orlando now undergoing an upgrade that is very challenging to drive through. Lane markings shift back and forth for miles and disappear at times. Lanes come and go. The roadway has sections where the surface is very uneven. I'm just happy that I don't have to drive through there very often.

With both AP1 and AP2, if the vehicle in front of me changes lanes to the right or left, my car using TACC brakes suddenly for no apparent reason. This is an invitation to be hit from behind by an inattentive tailgater. I find that I can overcome this by being alert and using the accelerator to keep moving.

The SW will get better but AP2+ has been around for over a year and a half and I detect very little progress on fundament functions. The cars do parallel parking splendidly, though.

carlk | 27/08/2018

TeslaTap Either way those who are betting on LIDAR have lost the lead of achieving general FSD. The main point for Tesla to select camera system is it could be put in every car now. Even if Waymo and others wait for LIDAR price to come down it will be years before they could put it in hundreds of thousands of cars to collect the machine learning data that Tesla has already been doing for a while. The argument can even be made that those companies might be better off to abandon it now and switch to camera than to wait for low cost LIDAR to become a reality if that will ever happen. The main point is the "brain" is the part that needs more work now not the "eye".

SamO | 27/08/2018

@george,

It feels like you are moving the goalpost. "General purpose FSD" is just a group of words you put together but has no relation to the term full self-driving. If you mean SAE Level 5, no steering wheel, zero driver possible input, that's fine, but not what this thread is about.

Full Self Driving is already here. Please review the links.

Do you really think that a current Model S/X/3 couldn't do the exact same course that Navya takes in Las Vegas or Switzerland?

I do.

You keep picking a corner case like your gated community or your garage, and then you say "see, it can't handle it now" (and silently you say "and I just don't understand how it will be in the future."

What I keep coming back to is that Tesla doesn't have to master your gated area. It can geofence the offering to public property. Or private property that has 1000 manually driven trips. Or to the city limits of Santa Monica.

Anywhere that makes it legal to operate, Tesla will be able to operate their vehicles in the Tesla Network. They already have 10,000X the volume of miles as their next closest competitor, Waymo. And their data collection lead is only growing. By 8,000 collector robots per week.

Care to guess when the first locality (state/city/country) will allow Tesla FSD anywhere?

Yodrak. | 27/08/2018

"Full Self Driving is already here. Please review the links."

I've reviewed the links, and along with some others I disagree. To me, "full" doesn't mean 1/2 full or 3/4 full, or SAE levels of full - full means full, 100%. Any thing short of that is self-driving in limited circumstances, which is what we have today. It isn't "full".

SamO | 27/08/2018

Sorry, Yodrak. But legally, your definition is simply wrong. Vernacular speech isn't appropriate for technical discussions.

georgehawley.fl.us | 27/08/2018

@SamO: I ride the autonomous FSD train from the central terminal at Tampa airport to the satellite terminals all the time.
I get your limitation on "full"Self-driving. Full isn't full. Truth isn't truth:-)) I am reminded of this when TACC turns off in the morning when driving east on a sunny day. Whatever full is limited to, it will still not be 100% full. But I am so old I won't have to worry about the the distinction unless I find my way to that little town in Switzerland and find a driverless bus spinning its wheels in the snow. :-))

SamO | 27/08/2018

@george,

I rode the autonomous shuttle/FSD shuttle in Vegas. No operator intervenes on the entire route.

It's full, whether you can make it there or not.

carlk | 28/08/2018

Uber just announced it will move away from cars and getting into bikes and e-bikes for short distance transportation. What I'm reading from this new strategy is very significant. It looks that Uber does not see much possibility of it to develop autonomous vehicles to replace its fleet before others do. It will fight them with "user operated" shared vehicles instead. Good luck to them.

SamO | 28/08/2018

Where I live in Santa Monica, we have several companies (Bird and Lime) that have dockless scooters throughout the city. A $7 Uber/Lyft ride is $2 on a Bird scooter.

This "last-mile" solution works great in areas with a warm climate. Luckily, with global warming, everywhere is getting a warm climate.

:-(

carlk | 28/08/2018

But then again there are those extreme weather conditions that come with it. It's not all "good" for riding your bike.

Mike83 | 28/08/2018

I love riding my electric bike for fun but it is hard to drink a cappuccino or carry stuff and I don't like riding where there is traffic esp. diesel and gas fumes. I did see someone riding with a gas mask. The new US business might be selling bottled air to breathe. My SCUBA tank is too heavy to carry around ;-)

RadOne | 28/08/2018

@Yodrak.
+1.
If FSD works only in geographically constrained areas, why would anyone outside those areas purchase it? And will it only be limited to certain systems? Far away from real autonomous driving.

bp | 28/08/2018

It seems more likely FSD regulatory approval will come in stages - since the government and politicians will be reluctant to approve anything that could increase risk of injury or death.

Even if the technology could run unattended, without regulatory approval and coverage by insurance companies, FSD will only be able to operate in "driver assist" mode.

So what seems likely is limited approval initially, under specific conditions - urban driving, requiring the FSD systems to drive slowly and/or stop whenever there is a risk of harm or on limited access highways or HOV lanes.

Getting the technology working is only the first step - then we hit the two L's - laws and liability - and that could take a while to resolve...

Yodrak. | 28/08/2018

"legally, your definition is simply wrong. Vernacular speech isn't appropriate for technical discussions."

Make up your mind - is this a legal discussion or a technical discussion? Or neither? In any case, I am not writing in the vernacular.

carlk | 28/08/2018

The driving force is still technology. When technology is advanced to a point that machine is (much) safer than human driver there is no way regulation will be allowed to prevent its implementation.

RadOne | 29/08/2018

I would like to see an autonomous car driving in all situations; highway, side streets, ball bouncing out in the street, child coming out between cars, unmarked streets, construction zones, those guys holding up the stop/slow signs in a one-lane work area, tree down in the road, unmarked grass parking lot, completely filled parking lot at Costco, etc. A car in a carefully maintained, limited zone is not anywhere close to be where truly autonomous vehicles have to be. Relatively just a high school science project now. Sure there are still tragic accidents, but when will technology be able to handle these situations, new situations and multiple variations better than a human driver?

SamO | 29/08/2018

Soon.

RadOne | 29/08/2018

@SamO. Depends on your definition of soon. Long way to go. Can't even recognize a stop sign yet or discern between a shadow and a real object. At least a decade to level 5.

RedShift | 29/08/2018

+1 RadOne

Having real knowledge of the hardware and software challenges, and being closely related to the founder of a self driving hardware startup currently raising money, I have to agree that level 5 is a decade away.

What are the current hurdles? Many: Speed of response (yes, even with very fast processing elements, this is a major issue today), redundancy, power consumption, cost, inadequate sensing technologies, and finally - the gigantic set of permutations and combination of real world situations, leading to a very large verification space.

carlk | 29/08/2018

Have a little faith in the capability of AI. We already have, or soon will have, powerful enough vision system and processor to do the job. It's only matter of time for us to teach the machine to learn. Even Lewis Hamilton or Walter Rohrl were clueless when they first sat behind the steering wheel. Machines can learn even faster than how we learn. And we can always put a better "brain" in the machine when needs to.

This recent event of AI vs human player in video game is a good illustration. Even that human gamers win the contest AI has proved its potential. Unlike Go or poker games those complicated video games need great speed to process info and hand to eye coordination. It's arguably way harder than driving a car. That machine could even get close to best gamers tells you how much potential it has.

https://www.theverge.com/2018/8/28/17787610/openai-dota-2-bots-ai-lost-i...

Basically Elon is thinking by the first principle, everything fundamentally needed to make it to work is there, while people who have doubts can only think by analogy.

carlk | 29/08/2018

You can see the machine is indeed learning pretty fast.

https://www.youtube.com/watch?time_continue=1&v=XX7M7K4QFQk

georgehawley.fl.us | 29/08/2018

@carlk: sorry to report that the software engineers are learning over a period of months. The car isn't learning anything yet but it will someday. I think it was ironic early in the video that a squirrel appeared in the road. The driver intervened saying that he was alert. The car didn't have a clue.

One of the trickiest issues is driving up a hill around a curve. The lane markings disappear over the horizon created by the top of the hill. The software needs to recognize this challenge and slow down as it approaches the top of the hill in order to better gauge the sharpness of the turn. It appears to do this which is great.

My poor old AP1 MX goes across double yellow if a curve is too sharp and it will never get any better since Mobileye went bye bye. :-((

SamO | 29/08/2018

@RedShift,

Tesla's competitors NEED full self-driving cars to be 10 years away. You (and the rest of the field) are going to find out the price of not having a product when the market demands one. Ridiculous valuations right now based on VCs not understanding the technology.

Teslas CAN recognize stop signs. Watch the autopilot video or any of the owners who got root access to all of the cameras and data streams.

In addition, their billions of miles of data and their lead over their competitors only gets bigger with every passing day. Look at Waymo "the leader" in FSD . . . they have 5M miles of data. They could buy 20,000 i-Pace and meanwhile, Tesla sold another 8,000 data collecting robots this week.

That don't need lidar to work.

RedShift | 29/08/2018

SamO

I’m not competing with Tesla on self driving hardware. Really, you presume too much.

RadOne | 29/08/2018

Billions of miles won't help in the everyday changing circumstances that occur. New construction, one-lane stop/slow road work, detours, accidents, obstructions, etc. These require more than just a mouse on a track. That is the difficult part.

carlk | 29/08/2018

@georgehawley
***sorry to report that the software engineers are learning over a period of months. ***

Machines are learning by themselves without software engineers involved in the improvement process. Read the article maybe you can get some more understanding of how AI neural network works. It's basically trail and error repeated billions and billions of times. Even programmers do not know how results are arrived. Their only control is to set up the rule and reward system and let the machine to run free after that.

@RadOne
How do you handle changing circumstances? If you can learn how to handle that machines can too.

Remnant | 30/08/2018

@bp (August 27, 2018)

<< ... Tesla's planning to upgrade the AP processor next year. >>

I didn't know that.

It certainly underscores Elon's dedication to the matter, as well as his ability to respond to the challenge appropriately and consistent with its complexity, even though it implies that the initial HW release may have been a tad premature (which many suspected from the get-go).

Al1 | 30/08/2018

@RadOne
How do you handle changing circumstances? If you can learn how to handle that machines can too.

Machines can also learn for you from the ones who were there before you. Circumstances do not change by themselves. Constructions are planned. The roads we have today, they were built for cars. There is no reason to believe humans won't be able to build infrastructure supporting FSD.

Remnant | 31/08/2018

@Al1 (August 30, 2018)

<< There is no reason to believe humans won't be able to build infrastructure supporting FSD. >>

There might be though.

Check this article in The Guardian, on "Franken-algorithms: the deadly consequences of unpredictable code"

It seems algorithms can acquire a life of their own and execute unexpected and unpredictable instructions that might result in damage, even death.

https://www.theguardian.com/technology
/2018/aug/29/coding-algorithms-frankenalgos-program-danger

carlk | 01/09/2018

That's the real danger of AI. Programmers can only set rules for reward and punish but there is no way for them to trace how results were arrived like they could with regular software programs. Even if you set the rule to have machines to only do right things one day they may decide the right thing is not to have any human around.

Al1 | 01/09/2018

Oh, I was merely referring to tell AI about new construction zones and suggest alternative roads.

Al1 | 01/09/2018

I understand though there might be risks.

Humans can be quite self destructive too.

carlk | 01/09/2018

***Humans can be quite self destructive too.***

Absolutely. The dangerous part is AI neural network pretty much "learn" the same way as we do. We still don't know fully why people do what they do. Mass murderers and serial killers grow up in the same environments everyone else does. And you usually could not detect a sign when they were little or even up to when they have committed the crime.

Now we only need to see which one will win the competition of destroying the humankind. I still think machines have the upper hand.

NKYTA | 01/09/2018

All the more reason to populate Mars, and Apha Ceturii II. Soon.

carlk | 01/09/2018

If we can't be space creature maybe machines can. That will at least make us the god so it's not all bad.

SamO | 28/09/2018

https://www.bloomberg.com/amp/news/articles/2018-09-28/tesla-enlists-emp...

Elon Musk has asked for Tesla Inc. employees to test what the company has billed as full self-driving capability and is dangling $13,000 in savings to entice them to help.

Musk wrote in an email obtained by Bloomberg News that Tesla needed about 100 more employees to join an internal testing program linked to rolling out the full self-driving capability. Any worker who buys a Tesla and agrees to share 300 to 400 hours of driving feedback with the company’s Autopilot team by the end of next year won’t have to pay for full self-driving -- an $8,000 saving -- or for a premium interior, which normally costs $5,000, Musk wrote.

JustSaying | 30/09/2018

"Elon Musk has asked for Tesla Inc. employees to test what the company has billed as full self-driving capability and is dangling $13,000 in savings to entice them to help"
Does that indicate that the current onboard computer processing power is Adequate for FSD?

TeslaTap.com | 30/09/2018

If it's only offered to a 100 employees, I expect these employee cars will be outfitted with the new HW3, well before the production rollout. This is a reasonable quantity to build pre-production units, and it's know that Tesla has had some cars equipped with HW3 internally for at least a month.