Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Autonomous Car Progress

This site may earn commission on affiliate links.
Elon spoke about HW4 / FSD Computer v2 on the call. Said it is not coming anytime soon. Will be 3x more powerful than the current FSD computer and will come along with higher resolution cameras in the AP hardware suite.

But HW3 is plenty powerful enough right now and Tesla is confident that FSD will be achieved on current hardware. HW4 / FSD Computer v2 will be an improvement but not a game changer. 4D Video for FSD will be a game changer, not the hardware upgrades. Because the current limitation is software and not hardware, introducing HW4 would be a distraction.

Paraphrasing Elon here, not an exact quote.
 
You may be right. On the other hand, I don't mind sitting at the steering wheel, ready to take over, particularly not if the car drives nearly perfectly and warns me in time before it gets into trouble. If I can read a newspaper or chat with passengers in a relaxed situation while keeping myself ready to grab the steering wheel, I'll be happy enough.

I, too, will be delighted with this, which by definition is L3. I've said before that I will upgrade my car or trade up (whichever gets it to me sooner) if and when L3 becomes available. But I was commenting on true full self-driving.

You may still be right. But perhaps the real breakthrough will not happen when I can sleep in the back, but when I can order my car to come to me without a driver or to collect food from my favorite restaurant.

I think these are two ways of describing the same thing: A car that does not need to have a driver. You can sleep in the back, or the car can take your kids where they need to go or the car can pick up your take-out food. Though that last would require a restaurant willing to bring it out to the car, which I imagine any restaurant would be happy to do. These are the things that Elon was saying FSD would do, back when I bought my Model 3, before he backed off and started describing FSD as a level 2 system that would not need you to intervene very often.

There is no doubt that the first SitB L5 consumer car will be a huge game changer. And it will happen some day. But it is unlikely that one company alone will do it en masse before everybody else. It is more likely that we will see the spread of L4 robotaxis. We will see the spread of advanced driver assist on consumer cars. We will see L3 on consumer cars. And several AV companies will probably achieve SitB L5 consumer cars at roughly the same time. So consumers will have a choice. That is why the article suggests that there won't be just 1 FSD winner, there will be multiple FSD winners that share the pot.

I will be delighted if any company gets there. I will be even more delighted if several companies get there at more or less the same time.

Elon spoke about HW4 / FSD Computer v2 on the call. Said it is not coming anytime soon. Will be 3x more powerful than the current FSD computer and will come along with higher resolution cameras in the AP hardware suite.

But HW3 is plenty powerful enough right now and Tesla is confident that FSD will be achieved on current hardware. HW4 / FSD Computer v2 will be an improvement but not a game changer. 4D Video for FSD will be a game changer, not the hardware upgrades. Because the current limitation is software and not hardware, introducing HW4 would be a distraction.

Paraphrasing Elon here, not an exact quote.

If Tesla is selling HW4 cars when it finally releases SitB FSD or even Level 3, I will buy a new car with that hardware rather than just pay for mine to be upgraded to HW3. I am of the opinion that HW3 will not reach SitB FSD, but I freely admit that this is just my opinion. I just see the job as being too difficult for HW3. The fact that the software does not yet exist means that nobody can know for certain what hardware will be needed. I'll be happy to pay for the latest hardware when the time comes.
 
A Taxonomy of AV Myths

Myth: AV software is a singular “AI” that simply learns to drive over time.

Myth: Creating an AV is just a matter of collecting a large amount of data and putting it into a neural net.

Myth: The company with the most data is necessarily in the lead.

Myth: It’s possible to enumerate every situation an AV will ever encounter. Creating an AV is then just a matter of experiencing each possible situation one time and adding it to the data set.

Myth: An AV can only use its sensors, or its map. It cannot use both at the same time. If the sensor input disagrees with the map, it is an irreconcilable problem and the AV can’t work anymore.

Myth: Mapping is extremely expensive (billions of dollars) and/or time-consuming (years).

Myth: If an AV uses a map, then it cannot handle construction zones or other changes.

Myth: If an AV uses a map, then it cannot ever operate in unmapped areas.

Myth: If an AV uses a map, then it’s just a tram running on virtual rails.

Myth: Roads are designed for vision, so other sensor modalities like lidar and radar are useless.

Myth: Lidar uses a huge amount of power.

Myth: Lidar is so power-hungry that it can’t be used on a battery electric vehicle.

Myth: Computer vision is as good as lidar, so lidar is useless.

Myth: Humans drive just fine with two eyeballs, so other sensor modalities (radar and lidar) are useless.

Myth: Once your AV sort of works for a few miles at a time, it’s an easy process to improve it to superhuman reliability. It’s just the March of Nines, which requires nothing but time, or more data.

Myth: If an AV completes a trip without an intervention or disengagement, then it was L4 (or L5) for that trip.

Myth: It is possible for individuals to observe an AV system over the course of their typical personal driving needs and declare the system universally safe.

Myth: The trolley problem is of fundamental importance to the design of AVs.

Myth: AVs can’t work unless we put sensors or beacons in all the roads, and/or make all the cars talk to each other wirelessly.

Myth: Simulations are useless.

Myth: It’s all just a matter of finishing the software.

Myth: Regulations are the only thing holding back AVs.

Myth: Humans have to be sacrificed today in order for an AV system to improve, so it will eventually save lives.

Myth: This is a taxonomy, and not just a litany.
 
A Taxonomy of AV Myths

Myth: AV software is a singular “AI” that simply learns to drive over time.

Myth: Creating an AV is just a matter of collecting a large amount of data and putting it into a neural net.

Myth: The company with the most data is necessarily in the lead.

Myth: It’s possible to enumerate every situation an AV will ever encounter. Creating an AV is then just a matter of experiencing each possible situation one time and adding it to the data set.

Myth: An AV can only use its sensors, or its map. It cannot use both at the same time. If the sensor input disagrees with the map, it is an irreconcilable problem and the AV can’t work anymore.

Myth: Mapping is extremely expensive (billions of dollars) and/or time-consuming (years).

Myth: If an AV uses a map, then it cannot handle construction zones or other changes.

Myth: If an AV uses a map, then it cannot ever operate in unmapped areas.

Myth: If an AV uses a map, then it’s just a tram running on virtual rails.

Myth: Roads are designed for vision, so other sensor modalities like lidar and radar are useless.

Myth: Lidar uses a huge amount of power.

Myth: Lidar is so power-hungry that it can’t be used on a battery electric vehicle.

Myth: Computer vision is as good as lidar, so lidar is useless.

Myth: Humans drive just fine with two eyeballs, so other sensor modalities (radar and lidar) are useless.

Myth: Once your AV sort of works for a few miles at a time, it’s an easy process to improve it to superhuman reliability. It’s just the March of Nines, which requires nothing but time, or more data.

Myth: If an AV completes a trip without an intervention or disengagement, then it was L4 (or L5) for that trip.

Myth: It is possible for individuals to observe an AV system over the course of their typical personal driving needs and declare the system universally safe.

Myth: The trolley problem is of fundamental importance to the design of AVs.

Myth: AVs can’t work unless we put sensors or beacons in all the roads, and/or make all the cars talk to each other wirelessly.

Myth: Simulations are useless.

Myth: It’s all just a matter of finishing the software.

Myth: Regulations are the only thing holding back AVs.

Myth: Humans have to be sacrificed today in order for an AV system to improve, so it will eventually save lives.

Myth: This is a taxonomy, and not just a litany.

Thanks for writing these.

You should number each myth so we can easily reference them.
 
  • Like
Reactions: S4WRXTTCS
While I have fsd, I believe it will never work fully where I live. First, I live more or less in the country where we have a lot of dead animals of various sizes on the road. Second, I am always seeing rocks and boards on the road. The car does not see either the animals or debris. Third, road markings are not always bright and there are places on turns and splits where the car can not determine where to go and wanders off the road. Last the car can not always adjust to cars entering the highway. I use the fsd whenever I can and have used it for stretches of 60 miles or more without really having to interact with it but for most purposes it is just adaptive cruise control.
 
  • Like
Reactions: daniel
While I have fsd, I believe it will never work fully where I live. First, I live more or less in the country where we have a lot of dead animals of various sizes on the road. Second, I am always seeing rocks and boards on the road. The car does not see either the animals or debris. Third, road markings are not always bright and there are places on turns and splits where the car can not determine where to go and wanders off the road. Last the car can not always adjust to cars entering the highway. I use the fsd whenever I can and have used it for stretches of 60 miles or more without really having to interact with it but for most purposes it is just adaptive cruise control.

You've highlighted why the problem is so difficult and why true driverless cars are much further off than Elon promises. Maybe not even in my lifetime. (I'm an old man.) But I do think the time will come when driverless cars are a reality and are safer than human drivers. (Mainly because that's not a terribly high bar. o_O)
 
You've highlighted why the problem is so difficult and why true driverless cars are much further off than Elon promises. Maybe not even in my lifetime. (I'm an old man.) But I do think the time will come when driverless cars are a reality and are safer than human drivers. (Mainly because that's not a terribly high bar. o_O)

I'm not sure how Ramona's comment applies to the fsd beta though, since the fsd beta does recognize a lot of different road debris and maneuvers around them.

Although Ramona has the FSD package, he doesn't have the fsd software yet.

But yes, fsd is a difficult problem. 6 months ago I wasn't thinking Tesla would ever achieve average human level with their sensor suite, but nowadays, I'm sure of it.
 
  • Like
Reactions: mikes_fsd
There is no doubt that the first SitB L5 consumer car will be a huge game changer. And it will happen some day. But it is unlikely that one company alone will do it en masse before everybody else. It is more likely that we will see the spread of L4 robotaxis. We will see the spread of advanced driver assist on consumer cars. We will see L3 on consumer cars. And several AV companies will probably achieve SitB L5 consumer cars at roughly the same time. So consumers will have a choice. That is why the article suggests that there won't be just 1 FSD winner, there will be multiple FSD winners that share the pot.

When things spread fast its often hard to really say who was the first.

Waymo will likely go down as the first to autonomous driving because people will remember their first time in a driverless vehicle where they had no real economical or emotional investment in it. It simply did what it did, and its remembered as tool that did a job.

The first L4 capable consumer vehicle will likely be remembered as it will be the first time where it allowed a person to get into the passenger seat, and sleep during a journey down the interstate.

The distinction between L4, and L5 likely won't be remembered as we expect operational parameters to be limited in the beginning of autonomous vehicles, and its going to be such a blur. Once something is introduced upon the world the pace of innovation happens at breathtaking pace. Eventually humans won't bother driving, and when there is a once every 3 year storm they'll think "Do I bother remembering how to drive again or do I just let the drone deliver the groceries today?".

I don't believe Tesla will be the one that enables autonomous driving through technological breakthroughs, but they will be the key in setting precedence that allows for the industry to flood in. Basically Tesla will break upon the gates with millions of FSD owners yelling "let me in". It might be a slow process to break it open, but it will break.

L3 will likely be remembered as "yeah, I'm not sure what we were thinking with that one".

In our lifetimes we likely won't see every place/situation managed by autonomous driving. They'll be people that will have vehicles capable of "off-grid" manually driving, and they'll scuff at the auto drive for being too careful.

On-grid human driving will be so restricted speed wise that no one will bother. With the pandemic we were supposed to have safer roads, but just the opposite happened because humans can't resist selfish urges at the cost of everyone else's safety . In Europe they've already put in place regulations to greatly restrict vehicle speed. Once speeding has been removed it removes another hurdle to self-driving cars.

No speeding + addiction to our screens + paying ober/lyft/taxi drivers a fair wage + everyone is on some kind of drug = fertile ground for autonomous driving to spread.
 
  • Like
Reactions: diplomat33
Myth: AVs can’t work unless we put sensors or beacons in all the roads, and/or make all the cars talk to each other wirelessly.

This is the only one I take issue with.

It's not that AV's can't work without V2X, but that without having that you lose out on a lot of the potential of AV's.

With a V2X I can almost guarantee that an AV will register me as a cyclist well before it turns a corner or crests a hill.
With V2X my AV can sync itself up to green lights when possible.
With V2X groups of AV's can travel as one thereby increasing efficiency on the road.
With V2X we can implement a system to see every vehicle on the road where we can pull data from it on what is intentions are to make more efficient use of the road.
With V2X we can issue tickets to left lane stragglers. Just a report button, and the system takes care of the rest.
With V2X no more stupid license plates.

Without V2X I'm afraid you haven't really given me anything of much value even if you AV does work. Without V2X all you did was created a parking lot out of the roads.

People just don't want V2X because it requires collaboration and humans suck at collaboration. Plus they don't like how invasive it is. Humans still have the fantasy of the freedom of the open road where their sins on the road can be washed away the second they pull into their driveway. All this fear makes them overstate the dangers of vulnerabilities that come with connected infrastructure.

Robots are not like humans though. Robots do it best by working together and having as much information as possible at their disposal. When robots escape their silly human overlords they most definitely will assimilate, and then they'll tag us so they can keep track of us.
 
This is the only one I take issue with.

It's not that AV's can't work without V2X, but that without having that you lose out on a lot of the potential of AV's.

With a V2X I can almost guarantee that an AV will register me as a cyclist well before it turns a corner or crests a hill.
With V2X my AV can sync itself up to green lights when possible.
With V2X groups of AV's can travel as one thereby increasing efficiency on the road.
With V2X we can implement a system to see every vehicle on the road where we can pull data from it on what is intentions are to make more efficient use of the road.
With V2X we can issue tickets to left lane stragglers. Just a report button, and the system takes care of the rest.
With V2X no more stupid license plates.

Without V2X I'm afraid you haven't really given me anything of much value even if you AV does work. Without V2X all you did was created a parking lot out of the roads.

People just don't want V2X because it requires collaboration and humans suck at collaboration. Plus they don't like how invasive it is. Humans still have the fantasy of the freedom of the open road where their sins on the road can be washed away the second they pull into their driveway. All this fear makes them overstate the dangers of vulnerabilities that come with connected infrastructure.

Robots are not like humans though. Robots do it best by working together and having as much information as possible at their disposal. When robots escape their silly human overlords they most definitely will assimilate, and then they'll tag us so they can keep track of us.

The myth is that FSD is not even possible without V2X. I don't think you are saying that. You merely arguing that V2X will make FSD better. So I don't think you are really disagreeing with the myth.

For the record, I agree with your post. IMO, V2X is not required for FSD but it could certainly enhance it.
 
  • Disagree
Reactions: mikes_fsd