Perhaps a new approach to FSD is needed.

Perhaps a new approach to FSD is needed.

FSD would be most useful on repetitive routes, like the daily commute.

The way I understand the development now is one programming group identifies objects and another group chooses a path for the car to follow.

After many miles from the Tesla fleet the car will learn and get better and better.

I think things could be speeded up if it would let me participate in the training on the regular routes I take.

Disclaimer: I have EAP in my 2018 Model 3. No FSD yet.

On my drive through town, several of the streets are one way. Parking is allowed on the right only. Some treat it as a two lane one way street. There are no lane markings. So some treat it as a one lane one way street. There is a black line that goes irregularly down the middle of the street where tar has been applied to fill the crack. TACC goes crazy with this black line.

So how could I help. Give me a personal train option based on GPS. I say or enter “trip to the post office” “Train”. It remembers the speed I drive and the path I take. If I avoid a pot hole, it remembers that. On the one way streets on the route, I will drive on the left hand side if the street away from the parked cars. It will remember that. I go 30 mph. It will remember that.

I still will expect the car to identify unplanned things like a child crossing the street or a car backing out of a drive way. If construction starts or ends on my route, I will request that I be allowed to “Retrain “.

If Tesla wants to use my trained route for others that is OK, but others may prefer their own training.

Once I have a trained post office route stored, it should take me to the post office perfectly each time. If bad weather happens, I will drive myself and not use the stored trained route.

-TheJohn- | 17 July, 2019

On the Pro side I love the idea of being able to introduce Local Knowledge type stuffs to the neural net.
On the Con, much of this seems like geo fencing which is not so great since things change and rapidly at times.
Cool and different idea though. Might also run into the demonstrable fact vs opinion problem as well but enough reports filter noise.

gmr6415 | 17 July, 2019

@cquail, If you haven't watched it you should watch the autonomy day event to better understand Tesla's approach to FSD.

"The way I understand the development now is one programming group identifies objects and another group chooses a path for the car to follow."

That isn't how it works.

Techy James | 17 July, 2019

@cquail If you don't have FSD, then why would there be option for it to remember your drive. Got to Pay to Play with FSD when it's available. Also there is additional items, like aware of traffic lights and determining if it's safe to pull out. Since we know these items will require AP Hardware 3.0 to be able to do, your car without you paying for FSD wouldn't have the 3.0 version of hardware.
The other downside to this approach which is called geo fencing and @-TheJohn pointed out is approach Waymo and some other companies are testing, and this has limitation, first road construction sign, pothole, or detour due to natural/man made event and the geo fencing fails. This might work for commercial transport systems between 2 set points, but not really practical for personal cases.
In my case I have a total of easily 5 routes I take to work and back. The route I take varies based on time I leave, whim to make it unpredictable, or traffic. Related to the geo fencing issue, in my neighborhood there is occasionally cars parked on side of street requiring me to move into opposite lane to get around. Obviously I can always drive in opposite lane due to safety issues like car coming, so geo fencing wouldn't be a good choice again. We are back to the real world where vision is much better.

carlk | 17 July, 2019

Add to what others said this is not how Tesla train its neural net. Unlike everyone else's Tesla's system does not follow a pre-mapped route. It trains the system to drive, just like human drivers do, by looking at the general landscape and features to decide how to go. It will need to learn everything everyone would see anywhere in the country/world. That's why it needs tremendous amount of data and takes so long to do.

This is the only way to achieve the general autonomy. What others are doing is kind of like how a blind person gets around. He can memorize everything around the house or places he frequents, and with the help of a cane (Lidar) to do (almost) what a normal person could. He has no way to go further than that though.

richtrav | 18 July, 2019

I totally understand where cquail is coming from and have thought of similar approaches too. But instead of changing FSD development why can't Tesla use some of these simpler approaches to act as a kind of a stopgap until a mature FSD system can be safely employed? Let's be real, FSD is probably going to take way longer than Elon's optimistic timelines. If Tesla hasn't gotten past NOA by the end of the year it might behoove them to look at adopting some simpler options to tide people over for a while. Tesla is asking real money for this autonomy stuff and the grumbling will only increase as the delays mount.

I don't see any reason why your car can't keep a GPS log of the routes of your choosing like cquail describes and you could "train" it by doing something like dropping a pin on your car's map wherever there's a stop sign or traffic light or pot hole or a speed limit sign is missing. Call it geofencing or whatever you like, at least it's mimicking FSD, the difference would be that you would usually be the "brain", giving the car instructions on actions to take if it can't reliably do them itself.

For instance, as an example given above: you're nearing a traffic light that's been premarked on your route, and instead of making the car determine that your light is red or green and whether it's safe to pull out the car would prompt you on what to do as you approach the intersection or make you take over temporarily. If you give no input it would disengage AP. Other things like speed bumps or potholes are easier, the car would just know to slow down. You'll still need to be paying attention to the road in case of unanticipated events but that is going to be the case even when the first iterations of FSD are deployed.

As FSD improves these simpler systems would be phased out when their functions can be better performed by the computer, but in the meantime it would throw a bone to those who dropped thousands while giving Tesla some breathing room to get FSD right before they turn it loose on the public. If Elon keeps upping the price on FSD people aren't going to bite based on nebulous promises (or threats), they'll at least want to see some usable benefits.

cquail | 18 July, 2019

@ richtrav. You nailed it. FSD will get better, but it needs to take small and successful steps first.

You are driving and hear a siren. You wonder where it is coming from and prepare to pull over. What will a FSD Tesla do?

SnowZA | 18 July, 2019

@cquail Depends. Like the whole FSD thing, I don't think it is an easy problem to solve. Does the Tesla even have a sensor capable of hearing the siren? I don't think it does... Which means that it would have to react to visual cues. On the other hand, visuals should be good enough to handle that in most cases - the siren is basically there to draw people's attention so that they can look for the source of the siren. The car should always be looking... It should be able to see the flashing lights and react appropriately. Also, hearing a siren should alert the driver to pay attention and be ready to override the car if it doesn't react appropriately...
In an alternative scenario, you are driving and have the radio turned up. You don't hear the siren until the source is pretty much right on top of you. In that case, it's no different than the autopilot - you would be reacting purely on visual cues anyway.
The only places where humans might have a significant advantage over the FSD Tesla is the case where there is a siren coming down a side street where the view is blocked by buildings or something. The driver might hear the siren and be more careful, or might not hear it and proceed as normal. In the real world, in my experience, many people would hear it and just ignore it anyway...

SnowZA | 18 July, 2019

@Fishev "Which really goes into how infrastructure would need to change. Emergency vehicles would have to send out radio signal that the self driving vehicles would have to be able to process."

Why send out a radio signal? You already have an audio signal. Doing the radio signal would mean that you would have to add another sensor to the car to process emergency broadcast radio signals, AND you would have to add the transmitter to all emergency and police vehicles. The audio signal is not going away - it's there as much for pedestrians as anything else. If you have to build a sensor in for that anyway, why not just add an audio sensor to just take advantage of what is already there?

On the other hand, if you are going to build transmitters into emergency vehicles, why stop at just a warning that there is an emergency vehicle in the area? Why not make it transmit its GPS location and heading (and potentially next few turns as well)... Then if the sensors are built in to FSD cars, they would be able to evaluate that data, determine if they are in the way, and move appropriately...

gmr6415 | 18 July, 2019

There seems to be a lot of "should do" here implying the car has AI capabilities itself to perform those "should do" requests.

As noted in the autonomy day event the neural net collects data from the cars that's fed to AI computers, which later results in updates that are sent to the car. I don't think the car's hardware itself is capable of the AI that would be required to do the "should do".

SnowZA | 18 July, 2019

@gmr6415 Sure, the car's AI isn't currently capable of figuring out these situations on its own. That's kind of the point of collecting the data - it gives them the stuff they need to refine how the car will react in given situations. It isn't as simple as programming something like "if you see a pothole, dodge it".

Humans have incredible visual processing, and years of experience built up to help them figure out what it is they're actually seeing, and what to do about it. Even so, humans make mistakes. Cars don't really have that experience to rely on to figure out what they are seeing, or what to do about it, and certainly don't have the flexibility of a human brain.

The actual decision of what to do in most situations is quite easy. In this case, it should safely dodge the pothole. The problem is figuring out a) how to reliably detect a pothole, b) how to reliably detect everything around it so that it can figure out what is and isn't safe. The more data that is fed into the system, the more they can refine how the car recognizes these situations. Those improved detection algorithms are what is being sent back to the car, so that next time it encounters a particular situation, the AI more likely to recognize it as such and know how to react.

Joshan | 18 July, 2019

Apply for a job at Tesla is you think you can do it better.

lordmiller | 18 July, 2019

@Joshan - Everyone is an expert. Let the professionals do their work, or write your own self-driving software and show Tesla how it's done.

howard | 18 July, 2019

I get the perceived disadvantage to geo-fencing but it is just another data set that provides a more accurate base location. Waymo still uses all its sensors in Tesla fashion to handle the real-world issues. Having high-resolution mapping/location information just helps with accuracy. For instance, my most recent experience with NoAP just does not make any sense. Why can it not detect off ramps and on ramps betters? On my drive back from the Airport when the white line on the right disappears due to an off/on ramp the car drifts to the right trying to center itself. Then when the line re-appears after the off/on ramp it abruptly moves back to the left. This to me would be a fairly fundamental lane trackign function yet it does not do it. This is the same with AP since I bought the car nearly a year ago along with phantom breaking and lane centering drift. Detailed mapping would allow the car to know where the off/on ramps are, predict overpass shadows, etc. and not react to them. FSD has got a long ways to go. HW3 will help but I have not seen much improvement in the basic lane tracking so far.

I just got back from a week in Gilbert AZ and watched a lot of Waymos driving and was quite impressed with what I saw. Tuning left from Gilbert road onto the 202 Westbound the Waymo was in the inside turn lane just about a car ahead of me. The light turned green and off we all went. It matched speed with the turning traffic and merged into heavy traffic at full speed smoothly. Watched a few other Waymos in city traffic looking like they were being driven by real drivers. They are making it work really well. I would not think driverless is very far off.

rwa | 18 July, 2019

I don't think this is a productive road to go down. It adds tremendous complexity, computation, and storage requirements, and could easily make things worse rather than better.

You mentioned training for a specific pothole. The problem with this is potholes are dynamic. They get fixed, and grow in new places. Training for a specific pothole, and not potholes generally, is "overfitting" and is the bane of good machine learning techniques.

Joshan | 18 July, 2019

These people do no understand AI and Machine Learning and are talking about how you would code things 20 years ago.

dsvick | 18 July, 2019

@richtrav and @cquail, first, if you already have AP, EAP, or FSD then Tesla is likely already getting data from you and your routes.
The problem with your proposal is that what you're both describing isn't FSD and would require them to start over with a completely different programming problem. And once it's done it only works for your car on your routes, under specific conditions, and assuming no changes have been made. The upshot is that it only benefits you, while Tesla's method, while it may take longer will benefit everyone, in all conditions, on all routes.
They are working on completely self driving cars that can follow any route, in any conditions, at any time, not Roombas who can vacuum your living room by bouncing off of things.

dsvick | 18 July, 2019

@FISHEV - " Tesla is at a technological disadvantage despite the huffing and puffing from Musk. What you see working well are what Musk says wont work while Tesla doesn't have it at all."

You're making assumptions that have yet to be proven. You should preface those with "I think .... "

Just because everyone else is doing something one way doesn't mean that they're right or that both can function equally well. And, just because you can see some cars maneuvering about without driver input doesn't mean that they are any further along that Tesla in developing FSD.

Joshan | 18 July, 2019

Gotta love a random scrub on the internet giving business advice to Elon Musk. How about your post your business accomplishment that qualify you to be giving advice to Elon, Fish?

Here is Elons for reference:

In December 2016, he was ranked 21st on the Forbes list of The World's Most Powerful People. He has a net worth of $22.3 billion and is listed by Forbes as the 40th-richest person in the world.

I am sure he is waiting with bated breath and your further assessment of his work!

howard | 18 July, 2019

dsvick, just because you can see some cars maneuvering about without driver input doesn't mean that they are any further along that Tesla in developing FSD.

This seems to be a constant theme with absolutely no basis. Tesla does not have anything on the road that remotely resembles the capabilities of Waymo. Of course, other cars are further along with FSD than Tesla!

Perhaps you should preface your comments with "I think Tesla has the ability to get there someday". This is just way old. | 18 July, 2019

@howard - Then again Waymo has nothing that can go over 40 mph. Hundreds of thousands of Teslas can drive on the freeway at 70 mph with almost FSD today.

Now Waymo has a great product - except you can't buy it, it costs $100K or so beyond the cost of the $25K vehicle, does not appear to work in anything other than a clear day, is super ugly, requires a trunk full of electronics, sensors are damaged if you touch them, and it can't be mass produced. Now if you consider than ahead of Tesla, please attempt to buy one and tell us how it works out.

howard | 18 July, 2019,

More miss information! Waymo most certainly now goes over 40 miles an hour. I followed one on the 202 west bound merging at over 60 with a long line of cars entering the highway.

FSD IS NOT CURRENTLY AVAILABLE ON A TESLA. PERIOD!!! You are referencing AP/NoAP and that is a long ways off from FSD. This is just a joke.

We are talking about Full Self Driving capabilities, not costs, looks, etc. or any of the other sidebars that you guys like to bring up as deflection. Let's just stick to FSD shall we?

Joshan | 18 July, 2019
Joshan | 18 July, 2019

Actually teslatap posted facts and you posted you opinion howard...

howard | 18 July, 2019

Joshan, There is really no way to comment when you are apparently this clueless.

Joshan | 18 July, 2019

I agree compeltely and feel the same way about your posts.... | 18 July, 2019

@howard - It's always possible Waymo has FSD on the freeway, but it's never appeared around here (I live practically next to Waymo). I've never seen a Waymo car on the freeway here or have any friends. It's likely what you saw was the driver actually driving the car, not running in FSD. To give Waymo credit, it's dealing with difficult local street conditions. At lower speeds, the processors and sensors do not have to work as hard. So low speed is easier to work with, but the environment is harder than limited access freeways.

Also I never said FSD is available on the Tesla, but near FSD is right now. I drive it most days.

This is a talk about FSD approaches. A system that adds $100K to the cost of a vehicle may function great, but clearly not practical for 99.9% of car owners. The other limitations make the current Waymo solution impractical, but useful as a test system. Now can Waymo solve all the issues and get the costs down? Perhaps - but not today and likely not for years.

I'm sure Tesla is running similar FSD tests with their own test vehicles, but it's not as obvious as Waymo, since the cars look the same as the other hundred thousand Teslas on the road.

Magic 8 Ball | 18 July, 2019

@howard When you saw the Waymo on the highway were you able to verify if it was in service performing driverless cab duty?

howard | 18 July, 2019

I was a car behind and to the right. There were other people in the van and it did perform well so perhaps you are correct. No way for me to know for sure. I am pretty certain I saw one northbound on the 10 as well. I should have overtaken it on the highway to confirm. Next time for sure. I'll be back in AZ next weekend.

FSD as defined by Elon is level 5. The Tesla requirement to have constant driver engagement (hand on the steering wheel) would be SAE level 2 (AP & NoAP). Waymo appears to be at level 4/5 with the safety engineer in the vehicle but not required to be active at the controls.

Magic 8 Ball | 18 July, 2019

Yup, more internet misinformation from "not sure" @howard, not surprised.

howard | 18 July, 2019

It was a good question I was honest.

Magic 8 Ball | 18 July, 2019

Yet, you jumped to a conclusion without any critical thinking and tried to spread misinformation, it is what you do.

howard | 18 July, 2019

You are so very correct. Kudos to your astute observations and calling out all liars. You are the best. Please keep up the good work as always. Sure Tesla will be sending more accolades your way as well.

dsvick | 18 July, 2019

@howard - "Tesla does not have anything on the road that remotely resembles the capabilities of Waymo..."

These two videos, one from almost three years ago, say otherwise...

ReD eXiLe ms us | 18 July, 2019

No, howard. Anytime a safety driver or observing engineer is required in the driving position, the vehicle is not above Level 3 Autonomy.

Level 3 -- An Automated Driving System (ADS) on the vehicle can itself perform all aspects of the driving task under some circumstances.  In those circumstances, the human driver must be ready to take back control at any time when the ADS requests the human driver to do so.  In all other circumstances, the human driver performs the driving task.

Level 4 -- An Automated Driving System (ADS) on the vehicle can itself perform all driving tasks and monitor the driving environment – essentially, do all the driving – in certain circumstances.  The human need not pay attention in those circumstances.

Level 5 -- An Automated Driving System (ADS) on the vehicle can do all the driving in all circumstances.  The human occupants are just passengers and need never be involved in driving.

NHTSA | Automated Vehicles for Safety | The Evolution of Automated Safety Technologies

M3BlueGeorgia | 18 July, 2019

Big difference between Waymo running tests on I-10 than their current service in a geo-fenced Phoenix suburb.

Current Waymo released service is at relatively low speed, in a well mapped area without much weather issues, and currently is reputedly unable to make left turns against traffic.
It is reasonable to presume they have written and are testing software to allow them to use highways.

Waymo have a very different approach to Tesla in FSD development, and very different hardware requirements. They are currently sheltered from commercial concerns by Google's deep pockets, but I'm not yet sure how they are going to successfully moniterize this significant investment.

dsvick | 18 July, 2019

@howard - "FSD as defined by Elon is level 5. The Tesla requirement to have constant driver engagement (hand on the steering wheel) would be SAE level 2 (AP & NoAP)."

And, as you accurately said, what Tesla has on the road right now available to consumers is not FSD and it isn't marketed as such. However, the videos I posted clearly show that Tesla does, in fact, have systems in place to be pretty close to level 5, probably in limited environments (like everyone else) but obviously more autonomous than AP/NoAP.

Regardless of who is where, my original post was taking offense at the statement that Tesla was at a technological disadvantage, when they clearly aren't and that the better technology has not been determined yet, and may never be.

howard | 18 July, 2019

ReD eXiLe ms us,

In November 2017, Waymo altered its Arizona testing by removing safety drivers in the driver position from their autonomous Chrysler Pacificas.[19] The cars were geofenced within a 100 square mile region surrounding Chandler, Arizona.[19] Waymo's early rider program members were the first to take rides using the new technology.[19] Waymo began testing its level 4 autonomous cars in Arizona

Also see video below.


I have seen the FSD Tesla demo but was this a hardcoded route for demo? If this is the case (FSD level 3) from 3 years ago then why has Tesla not really advanced? At this point I don't find it comfortable, if not safe, to use TACC, AP or NoAp. I have never ever said Tesla does not have the capability to get there. We are discussing what is current today.

Check out this one from the same time period.


The big difference is that Waymo has been performing at this level for the last 2 plus years. If Tesla was doing the same anywhere it would be big news and reported everywhere. This would not be something Elon would keep quiet.

howard | 18 July, 2019

M3BlueGeorgia, but I'm not yet sure how they are going to successfully moniterize this significant investment.

I would guess their goals are not much unlike Tesla's. Once their fully FSD level 5 it will "revolutionize" transportation as Elon as repeatedly stated for Tesla. One of my employees only Uber's. He has some high user deep discount plan. I know he would jump at taking a less expensive option and never bother with car ownership. A lot of folks are going to be all over this. In another 10 more years I doubt I will own one either.

Waymo and Jaguar have partnered to develop the world's first fully self-driving Jaguar I-PACE. Testing began in 2018.

In the next few years, we'll add up to 20,000 I-PACE vehicles to Waymo's fleet. That's enough to drive about a million trips in a typical day.

ReD eXiLe ms us | 18 July, 2019

howard: You are forgetting the part that reads " certain circumstances...", which in the case of WAYMO amounts to specific geofenced locations. Their vehicles are NOT capable of Level 4 operation anywhere else. Everywhere else their Level 3 and below systems WILL require a human driver. Tesla's goal is to not need a human driver anywhere, ever. Hence, the word 'FULL' preceding 'self driving' in the description of FSD.

In any condition? I hope not. I now live in a place where it rains, far too often for comfort, with thunderstorms, sleet, and hail at times, though thankfully snow is a rarity.

I think autonomous cars should be immobile during inclement weather conditions, like I believe drunk people shouldn't be allowed to drive. A hardheaded drunken Friend once cracked up his Mom's BMW when I walked a girl to her apartment not far from a bar. I specifically told all my other Pals NOT to let him drive. I should have held him down and took the keys.

howard | 18 July, 2019

ReD eXiLe ms us, I get it and we can split hairs all day. We are simply talking about "current" capabilities. I agree with your arguments but it does not change the fact that Waymo is currently capable of and is operating Level 4 and has been for some time. It will probably be Level 5 in the not too distant future within it GeoFensed area. But let's face it where is the greatest use going to be? In a large metro area like Phoenix. So who really cares if it does not go outside the geofenced area. It is a moot point and weak argument. They are doing what 95% of future RoboTaxis, including Tesla, are going to be doing. I don't need a taxi to take me from Phoenix to LA. Do I?

We can speculate when and to what extent Tesla will be Level 4 and eventually level 5 till the cows come home, but that's the point it is just "speculation". It is not now and I don't think it is going to be very soon. I do agree that it will most likely get there but it has been delayed and delayed and delayed. Has it not? It is not going to be the end of the year as Musk has stated. Do you think so?

If I remember right I think Waymo is testing inclement weather up in Michigan or up that direction.

kaushal | 18 July, 2019

"instead of changing FSD development why can't Tesla use some of these simpler approaches to act as a kind of a stopgap until a mature FSD system can be safely employed? "

If you can get people to for something years in advance with no hard/soft delivery date, why wouldn't you do it? It's brilliant actually.

Reminds me of those IT contracting firms at large govt organizations who don't produce anything of value, but, continue to get paid every month!

cquail | 18 July, 2019

I am not suggesting that Tesla is on the wrong track, but the neural network takes a long time to learn. I am suggesting that Tesla drivers help speed things up by training their own car for repetitive routes.

New FSD problem: A four lane road goes through town. A police car is in the far right on coming traffic lane with lights flashing. Behind the police car is a funeral procession. Cars going in the opposite direction pull over and stop out of respect. What does the FSD Tesla do?

ReD eXiLe ms us | 19 July, 2019

howard: It doesn't matter what I believe anymore than it matters what you do. What matters is what will be and when. I have more trust and faith in Elon Musk than I do WAYMO, CRUZE, or George Hotz.

cquail: No one wants their RoboTaxi to be trained by Sheriff Buford T Justice.

Joshan | 19 July, 2019

@howard the fact that Waymo is currently capable of and is operating Level 4 and has been for some time. It will probably be Level 5 in the not too distant future within it GeoFensed area.

Geofenced lvl 5 huh? I find it HYSTERICAL the guy who has been arguing with me for a week on the definition of FSD stated Geofenced lvl 5...

I will let you figure out why that is so funny.

RES IPSA | 19 July, 2019

I would like to see Tesla perfect the Level 2 and 3 "autonomous" features before we start to speculate about Level 4 or 5.

The current state of Navigate on Autopilot is a clear example of how far we still have to go to even perfect Level 3 autonomous driving features. Hopefully by the end of 2020, Tesla can perfect Level 3 on highways and city streets. At that point, we may have a better idea of when Level 4 can come into being.

cquail | 20 July, 2019

People make driving decision based on more than what they have seen on the road in the past.

This spring I29 between St Joesph MO and Council Bluffs IA was closed because of Missouri River flooding. I hear this on the news and decide on an alternate route. Will the FSD car drive to the “Road Closed” sign before it alters it’s route? Will it have enough charge to execute plan B?

ReD eXiLe ms us | 20 July, 2019

cquail: Ah. How was the radio announcer informed of the road closure? Were they contacted by phone by their third Cousin Felix who erects barriers for the County Sheriff? Or did they receive an alert transmitted via computer from the authorities? What makes you think those same alerts wouldn't be communicated to an autonomous car, the same way they are transmitted to mobile phones -- automatically...?

Y'know, just like Emergency Broadcast Warnings, Severe Weather Alerts, and Amber Alerts. Even my older phones receive such automated messages if powered up, even though they are not assigned to a commercial carrier anymore. The long arm of the FCC requires it. I figure the same thing is true of Tesla cars as well as vehicles equipped with OnStar which are basically mobile phones on wheels.