My take on Tesla's current position in the industry.

[ecarfan]

I hope Tesla has a more affordable SUV type car in the next 5 years or so

Yes Tesla will produce such a car, it is often referred to as the Model Y and it will be built on the Model 3 platform. See [Updated] Elon Musk pulling an all-nighter for "Secret Master Plan" Part 2 quote:
"Today, Tesla addresses two relatively small segments of premium sedans and SUVs. With the Model 3, a future compact SUV and a new kind of pickup truck, we plan to address most of the consumer market."

Guesses as to when the Model Y will go into production range from 2017 (most think that is very unlikely) to 2018, 2019 or later. Meaning we really have no idea. Elon just says in the "future".

 

[JeffK]

The driving logic can and should be thought of as a set of algorithms with the NNs being tools utilized by those algorithms.

That's just it, the driving logic is itself a neural network built utilizing the output of other networks such as for image recognition, pathfinding, etc and human driving data... If you check out any of the MobilEye or Nvidia talks they describe this exact thing. The only time you need human programmed logic is when training the network such as labeling images, or negative rewards such as for hitting an object, etc. The rest the network has to figure out on it's own using the driving data. This is the only way it can be dynamic. If driving policy/logic was algorithm based then we'd never see autonomous driving in our lifetimes because we'd never be able to come up with all the edge cases and situations.

The MobilEye video The Three Pillars of Autonomous Driving with Prof. Amnon Shashua actually shows a slide describing this at about the 13 min mark.

Driving logic was a set of algorithms using neural networks as tools back in the self driving demos in the 1980s and 1990s... it's just not sufficient for real world driving.

 

[zenmaster]

That's just it, the driving logic is itself a neural network built utilizing the output of other networks such as for image recognition, pathfinding, etc and human driving data... If you check out any of the MobilEye or Nvidia talks they describe this exact thing. The only time you need human programmed logic is when training the network such as labeling images, or negative rewards such as for hitting an object, etc. The rest the network has to figure out on it's own using the driving data. This is the only way it can be dynamic. If driving policy/logic was algorithm based then we'd never see autonomous driving in our lifetimes because we'd never be able to come up with all the edge cases and situations.

The MobilEye video The Three Pillars of Autonomous Driving with Prof. Amnon Shashua actually shows a slide describing this at about the 13 min mark.

Driving logic was a set of algorithms using neural networks as tools back in the self driving demos in the 1980s and 1990s... it's just not sufficient for real world driving.

Look, it is silly to say all tasks can handled by some the black-box neural net. Level 4 autonomy is the goal which has new requirements as they are identified and understood, and thus requires new design work and yes, new algorithms. It's questionable whether the existing collected data ("millions of miles") even has utility for the new level 4 tasks required! Even for "deep neural networks" you *must* design algorithms to properly train them. All I'm saying is that is where the value is as far as making the tech sophisticated enough to handle fully autonomous driving. My whole premise was that Tesla may *not* be a leader in having the most sophisticated tech in the industry to handle level 4 autonomy requirements.They will not have full autonomy in 2018 like Musk carelessly promised. Or even years afterward due to the lack of problem identification (which you need to know in order to design an effective neural net).

 

[JeffK]

lack of problem identification (which you need to know in order to design an effective neural net).

Even that's not always necessary, see unsupervised learning.

Besides the image identification and a few other aspects which are simply labeled by humans, the driving itself is learned from watching humans plus simulated trial and error using reinforcement learning.

The "algorithms" meaning the weights of the DNNs are generated by the machine learning process and go into the final product. Any algorithm used to manipulate images is not necessarily needed in the final product.

Here's a pretty decent write up on some common self-driving misconceptions
Top misconceptions of autonomous cars and self-driving vehicles | Driverless car market watch

Consider a teenage driver, you make them read a book on some laws and safe practices but that same book doesn't necessarily tell them what to do in every circumstance. There are only a few basic rules which are location/government dependent. After they get their license then they learn by doing and observing others. Teens make mistakes and either correct them or Darwin catches up to them. This make the teens left living better drivers. The more miles they drive the more situations they encounter and the more reinforcement they get and they become even better.

The teen uses his sensors (eyes, ears, fluid in the inner ears, and sense of touch) and his DNN (brain made up of smaller neural networks that specialize in certain tasks) to interpret the sensor data and respond accordingly. Besides reading that drivers manual and maybe getting yelled at by an adult a few times the human brain doesn't need to be "programmed" in the traditional sense. Is it silly that a teen can handle all driving tasks by his black-box neural net? At first yes, of course, it's a teen, but with time and experience that black-box neural net can become a decent, if not great, driver.

An artificial dnn is not nearly as advanced as the human brain so it needs a lot more real world data. Once it has this, then it's only a matter of showing it's statistically safer than a human and having regulators make a decision on the matter.

Tesla has already shown they are leaders in the Level 2 dept. compared to volvo, mercedes, etc. Google is not going to sell cars, Uber's self driving cars are not for sale either. There is no one selling cars in the level 4/5 space until next year when Volvo starts selling a few Drive PX 2 powered cars.

In the machine learning world those who collect the most data are in the best shape. The current best neural net designs for recognition are pretty well defined in academic circles. Tesla can even iterate through custom designs quickly with enough computing power.

In the end we will only truly know they are a leader in level 4/5 tech and/or DNNs when they release statistical data to the government late next year. They are, however, first to market with the full hardware suite in a production vehicle.

 

[3Victoria]

They will not have full autonomy in 2018 like Musk carelessly promised. Or even years afterward due to the lack of problem identification (which you need to know in order to design an effective neural net).

That's rather a ballsy prediction. You must have access to internal Tesla details to make such assured statements.

 

[JeffK]

That's rather a ballsy prediction. You must have access to internal Tesla details to make such assured statements.

Also as far as I know they never gave firm date on enabling level 5 autonomy. (at least not during the announcement call)

heck, they gave themselves to the end of next year just to do a public LA to NY demo. It'll really come down to regulators.

 

[ItsNotAboutTheMoney]

I think the M3 is the make it or break it point for Tesla. Even more so now with the release of L5 hardware / public disclosure of enabling software. I think it is so much so that if Tesla fails to get as many of these cars into customers hands with AP enabled in the next 2-3 years you'll end up seeing Tesla become the fall back go to company for any issues or regulation policy's once the Fed's catch up with the technology. Understandable, the mass public opinion needs to be overwhelmingly positive as well.

What I am getting at here is, at the moment the microscope is focused on Tesla, while the other auto manufactures sit back and watch how this bleeding edge technology is going to unfold. They get to watch not only how Tesla is solving the problem and develop a competing product in the shadows, but also watch the public / government response. One might speculate that the legacy auto manufactures position is to use Tesla as a burner company. Use them as a beta test, then pivot their legislative influence they have accumulated over the years to force them out of their own market.

Model 3 has always been make or break.

 

[3Victoria]

Also as far as I know they never gave firm date on enabling level 5 autonomy. (at least not during the announcement call)

heck, they gave themselves to the end of next year just to do a public LA to NY demo. It'll really come down to regulators.

It was your statements about "years later" and "lack of problem identification" I have the problem with. Nothing ventured, nothing gained. Tesla is pushing the envelope.

 

[ItsNotAboutTheMoney]

It's hard to know what Tesla's position is relative to other manufacturers. Tesla is simply aggressive in marketing and releasing driver assistance technology and has always been taking an approach based on a target of autonomous driving. Other manufacturers don't need to be or aren't as aggressive in releasing features so can appear to be further behind than they are.

At this point, Tesla has only begun to use the AP2 hardware, and there's no significant features in software. We'll know more where they are in their software over the next year. Obviously, there are positive signs that we'll be seeing very advanced driver assistance within 2 years, and that finally we'll have real-world deployment of significant numbers of vehicles that will allow us to get past bluster and hype on autonomy.

 

[ecarfan]

They will not have full autonomy in 2018 like Musk carelessly promised. Or even years afterward due to the lack of problem identification (which you need to know in order to design an effective neural net).

Elon did not "promise" that date for Full Self Driving Capability.

I would be interesting in knowing your professional qualifications in automotive engineering / autonomous vehicle design / and neural net programming that provide you with the expertise to make such definitive statements.

 

[3Victoria]

It doesn't really matter exactly hiw much functionality is acheived and the absolute timing, only the relative timing. Tesla is pushing the development of EV and solar, and moving this forward and earlier. Left to their own devices, the existing manufactures would proceed very slowly, as has been amply demonstrated over the last twenty years.

Even if level 5 does not arrive soon, it will arrive sooner with Tesla than without. Level 4 will be incrediably useful and functional cake by itself ... level 5 is the icing.

 

[shrspeedblade]

I think a lot of the disagreement comes down to people thinking Level 5 will have to be perfect, but the more I think about it it doesn't have to be.

It does have to be near perfect, due to the fact that the legal industry (especially in the U.S.) has been practically turned into an organized extortion scheme with the practitioners receiving the lion's share of the the benefit. If not for that merely very good would be acceptable.

Think of it this way, Tesla can almost certainly already outperform with AI the average teenage (certainly this!) or elderly driver who is at that borderline stage. We all encounter people like this on an almost daily basis, which why driving defensively is so important. Then you have the average group, for whom driving is to get from point A to point B and cars are an appliance. If statistically they can get above this group you could argue Autopilot/Level 5/whatever is safer, but we all know they'll be held to a higher standard.

When they can get to the point that only the safest of human drivers: those with lots of experience but still with quick reflexes who actually drive well within their envelope while being courteous, looking ahead, paying attention, sticking close to appropriate speed for conditions (different than speed limits folks!) then I WANT Level 5 to be legal for everybody else because it makes me safer even when I'm just out cruising in my Miata!

To paraphrase a common quote, don't let the desire for the perfect be the enemy of the good (or maybe even great!)

 

[zenmaster]

To paraphrase a common quote, don't let the desire for the perfect be the enemy of the good (or maybe even great!)

The vast majority of driving requires no advanced AI at all. What is "good enough" when a circumstance occurring maybe 10% of the time/distance, but nonetheless occurs, requires the car to make a decision when it simply cannot due to some design limitation? What do you think the consequences of "not perfect" would be in the real world? The way I look at it, it's not "level 4" if it ever requires a driver in any circumstance. It seems likely that the industry will start qualifying "level 4" with many asterisks. When some tech starts becoming indispensable in everyday life, people will demand governmental accommodation for its shortcomings however.

 

[Dan Detweiler]

The vast majority of driving requires no advanced AI at all. What is "good enough" when a circumstance occurring maybe 10% of the time/distance, but nonetheless occurs, requires the car to make a decision when it simply cannot due to some design limitation? What do you think the consequences of "not perfect" would be in the real world? The way I look at it, it's not "level 4" if it ever requires a driver in any circumstance. It seems likely that the industry will start qualifying "level 4" with many asterisks. When some tech starts becoming indispensable in everyday life, people will demand governmental accommodation for its shortcomings however.

Judging from statistics, the results in the "real world" would be at least twice as bad.

 

[LectrikPower]

Yes Tesla will produce such a car, it is often referred to as the Model Y and it will be built on the Model 3 platform. See [Updated] Elon Musk pulling an all-nighter for "Secret Master Plan" Part 2 quote:
"Today, Tesla addresses two relatively small segments of premium sedans and SUVs. With the Model 3, a future compact SUV and a new kind of pickup truck, we plan to address most of the consumer market."

Guesses as to when the Model Y will go into production range from 2017 (most think that is very unlikely) to 2018, 2019 or later. Meaning we really have no idea. Elon just says in the "future".

The Y probably won't be big enough for me. I'm hoping X prices come down or I'll just get a used X.

 

[Red Sage]

I've always felt that Perfect was the direct enemy of, "Meh. That's good enough. Nobody will notice anyway..."

Someone will notice. Then they will go out of their way to point it out to everyone else, even if it doesn't actually matter. Most streets in urban areas in the United States are constructed on a 90 degree grid system. Someone will find the curviest, hilliest, most heavily blind-cornered, off-camber, acute angle, hairpin, cul de sac infested neighborhood in all the land and happily demonstrate exactly how 'dangerous' an Autonomous car would be in that area. So Tesla will have to anticipate such an action and make sure the attempt to force a failure does not come to fruition by preparing their cars for that situation well ahead of time.

Best block? No be there. Second best block? Punch them in the mouth. Hard.

 

[Jayc]

Okay here is a scenario trivia for someone with a background in machine learning / autonomous driving to answer.

Imagine it is 2020, the government has passed the remaining bits of legislation allowing full autonomous driving. You have just had a very difficult day at work and you decide to let your level 5 autonomous car drive you home while you have a cat nap. The car takes the usual route but soon finds itself in a very long line of traffic with no exit to an alternative route. Traffic moves at a snails pace and eventually reach the "problem spot" where a driver driving his car himself without autonomous driving has crashed onto the traffic lights so the traffic lights are out and there is a policeman at the side of the road directing traffic. Imagine your car's connected neural networks are yet to be informed of this situation as you were amongst the very first to get there.

1. How should the level 5 autonomous car handle this situation with the sensors and reasoning capabilities it has ?

2. There has actually been a fatality at the scene so a section of the exit road has also been cordoned off, there is a small hand-written sign on the side of the road asking drivers to take a diversion avoiding a certain area. Explain briefly the sensory inputs and the sequence of reasoning that the autonomous driving algorithm will take.

 

[Dan Detweiler]

Okay here is a scenario trivia for someone with a background in machine learning / autonomous driving to answer.

Imagine it is 2020, the government has passed the remaining bits of legislation allowing full autonomous driving. You have just had a very difficult day at work and you decide to let your level 5 autonomous car drive you home while you have a cat nap. The car takes the usual route but soon finds itself in a very long line of traffic with no exit to an alternative route. Traffic moves at a snails pace and eventually reach the "problem spot" where a driver driving his car himself without autonomous driving has crashed onto the traffic lights so the traffic lights are out and there is a policeman at the side of the road directing traffic. Imagine your car's connected neural networks are yet to be informed of this situation as you were amongst the very first to get there.

1. How should the level 5 autonomous car handle this situation with the sensors and reasoning capabilities it has ?

2. There has actually been a fatality at the scene so a section of the exit road has also been cordoned off, there is a small hand-written sign on the side of the road asking drivers to take a diversion avoiding a certain area. Explain briefly the sensory inputs and the sequence of reasoning that the autonomous driving algorithm will take.

Your car wakes your lazy ass up and says "Hey, you can take it from here".

Seriously though, I believe that I read somewhere that they programing software to recognize hand signals, which I guess means that your car can recognize and return the one finger salute you get on the highway during rush hour traffic.

Dan

 

[JeffK]

a small hand-written sign on the side of the road

Seriously, hand-written?

I've personally never come across this situation nor would I think my future self driving car would either. Also, Google maps are pretty accurate as far as traffic is concerned as reported by your state's DOT.

1. Very simply, the car will avoid obstacles and try to get you to your destination. The google car recognizes bicyclist hand signals, police hand signals are kind of low hanging fruit. Many people ask about that so you know that'd be a corner case that's definitely covered.

2. There's no way there will be hand-written signs, such crap, there will be official signs and/or more likely a cop car in the way of the exit. The car should proceed to the next exit obviously and reroute. There's only one logical choice here, it's not complicated.

We will have to wait and see how it does in real life when it's enabled. I can say with certainty that by 2020 we'll have racked up tens of billions of miles. The likelihood that these events wouldn't have been encountered before would be extremely rare.

Now if there's a handwritten sign with balloons pointing toward the house where little Cindy Lou Who is having a birthday party I wouldn't expect the car to read or follow that sign... her parents better have supplied an address or GPS coords if it's way out in the country where maps and county records still suck to this day.

 

[McRat]

I think everybody tried to feed the horse in the wrong end of the animal.

The safest systems will be based on area vehicles communicating with each other. Think of it as distributed computing. Each car's autonomy engines will be amplified by surrounding vehicles. Sure, it will be limited to begin with. But in 10 years you'd have the best system possible. Your car will know things that it's own sensors cannot tell it. Things it cannot "see". Actions that haven't happened yet. It would be so smooth that in the future, stop lights will only be for allowing pedestrians to cross.

Thing how much faster you can process your surroundings if you instantly know where cars are, what speed, what direction, and what future path the cars will be traveling. Normally you'd have to process each video frame to determine this. If there are 20 cars in the cameras, every frame must be evaluated to define car objects in them and fill in the trajectories.

10 years from now, Homer Simpson is going to go DOH!!! We are building an internet of cars without TCP/IP.