For me, Tesla’s Full Self Driving capability is the most exciting technology breakthrough since the six moon landings in the 1960s and 70s. Being able to train a computer system to work on any road network the world, safer than the best human driver, at a lower cost and with higher safety is a serious challenge. But it is nevertheless something which visionaries such as Elon Musk (love him or hate him) see in our future and not that far away either, more on this later.
The Full Self Driving Challenge
Creating a technology such as FSD with all the other barriers to entry such as:
- High cost
- Low chance of success
- At a time when there is a serious amount of Fear, Uncertainty and Doubt (FUD) in the news
- Without similar systems to “copy”
- With minimal research and expertise
is one serious challenge.
Tesla, however are in the best possible position to create not just a working but economically viable solution to the problem. This year, Tesla are projecting production of a little under 2 million cars for the year. All of them already pre-installed with all the hardware required to run current and future versions of “FSD”. Tesla assure us that the current hardware combination is enough to make cars fully autonomous.
How will cars become fully autonomous?
With a simple “over the air” software update, which all of Tesla’s car range have, Tesla are able to turn on the new FSD capability for a one off or lower monthly fee.
Once FSD is “solved” and ready for a truly driverless experience, a tap in the Tesla app will be all that’s required to make your car fully autonomous.
How will FSD work?
FSD v12 is a fully AI powered driverless car technology. Where as v11 is available to the public in Northern America and not 100% vision driven, v12 is only available to a select few for testing purposes, mostly Tesla employees at this time but is trained with 100% vision and no statically coded parameters to tell it what to do.
In FSD v12 there are little to no hard coded parameters. For example, the car knows to stop at red lights, not because someone labeled a bunch of red lights and coded the car to stop at one. The solution at high level is far more simplistic, depending on how you look at it. The system was trained on millions of video feeds, from this, the car understands the environment like a human brain and knows that what a car should do at any particular moment or “frame”. There is no knowledge of traffic lights or lane lines. The system just intuitively knows what you should do due to seeing so many examples from it’s training data.
This sounds simple, why hasn’t anyone already solved it?
The question to ask is actually, who has access to millions or billions of miles of driving data, which it has legal access to collect and train a system with? Let’s not also forget that a significant investment in GPUs/ TPUs is required to perform the training. Tesla has said they will be spending around $2B USD on training chips this year, and will probably exceed this next year. Only a few companie shave that kind of cash, but not all of them have the data or engineering talent to replicate what Tesla has done. Not only this, but you need cars to run this software on. Which other automotive company meets all these other criteria? None, yet!
Is Full Self Driving solved?
My option on the mater is that FSD is practically solved now. The backend infrastructure, process and procedure for FSD appears to be well developed (After a number of rewrites, which is the cause of all the delays)
To complete the solution, an enormous amount of training data is required for training the system, issues need to be detected then better data identified and used for those failure scenarios to refine the system. Then repeat.
This means that for new countries like the UK, the system will need retraining with different video feeds. But this isn’t much of a problem for Tesla since all the hardware including the cameras and internet connectivity is already included in every car they sell. The main constraint right now seems to be the acquisition of enough training compute in the form of GPUs or TPUs (or even the production of Tesla’s internal training chip) when there is already a worldwide shortage of such hardware.
As you can see from the below snippet from the resent Tesla Shareholder Earnings call, Tesla are not messing around when it comes to scaling their amount of compute.
Why do you need some much data?
Unlike humans, FSD needs an enormous amount of data for training purposes. But there is more to it that this. The system needs to be much better than the average human and if you are training the system on based on what the average human does, how do you make it better? The answer is to detect times when the system does something wrong and locate higher quality training data for that particular scenario, then retrain.
In Elon’s recent FSD v12 livestream, the car tried to move forward at a red light. This was likely because there were so many red lights in the frame the system saw a green one and wanted to proceed. By pressing the brake pedal, turning off the system or making a different steering wheel movement, Tesla can capture the video feed for these moments, determine what the issue was and collect more data for this event for retraining. In order for the system to get better than the average human, Tesla must calculate who are better drivers or better and more specifically what video feeds are of better driving quality and use more of these for their training.
There are also all the edge cases to consider too, Potholes, flying garbage bags and wild animals roaming the streets! These situations are only solved if you have enough examples of edge cases. This should be no problem long term, given Tesla’s ever increasing number of FSD capable cars on the roads. Remember, even if you have not subscribed to FSD, the car can still anonymously upload video clips back the HQ for training purposes (providing you have given consent of course!)
So in summary, more data is necessary as you want to end up only using the highest quality examples of driving in the training set.
So when will FSD be available outside of North America?
Right now “FSD” is technically available in places like the UK and Europe, however it’s highly limited by regulation. For example you can buy “FSD” on any new Tesla for around half the price you can in the US. This is because it has around half or even less than the the capability of the US system.
The regular FSD in the UK and Europe offers auto lane changing, but only after the driver initiates it with a tap of the indicator. This is limited to “controlled highways” essentially roads where pedestrians are not allowed like Motorways and other high speed roads. Other features such as stop sign and stop light recognition are also included within the regular FSD bundle.
Dates for the newer AKA FSD Beta (v11 or even v12) have not been released yet, but there are some hints.
Elon has recently posted that he wants to see this new version of FSD working in other countries before they offer it to even the latest North American hardware 4 vehicles. This is good news for us!
Elon doubled down on this statement in a later reply by saying “a real 6 months”
International FSD in 6 months – But what about the regulators?
There has been some good news on this front too, according to various sources, the UNECE (Who are creating a framework for other countries to adopt for Autonomous Driving) are meeting more regularly than usual. This is probably due to an increase in public interest and how many injuries and fatalities even existing data shows will be reduced due to FSD type technologies being adopted on our roads:
Sources are telling us that it’s possible for more FSD style capabilities will be permitted within the framework very early in 2024. These will be “hands on” capabilities but nevertheless a giant step forward.
If the framework is completed in early 2024 and adopted by the member countries (UK, Europe, Japan etc) then Elon’s timescales of 6 months could be accurate!
The latest information I could find in relation to “hands off” approval is due for updates towards the end of 2024:
What might hold up Tesla’s International FSD rollout?
If we take the version of FSD Beta than those in Northern America enjoy today and assume it just works in the UK with minimal re-training, then there are still regulatory complications not just on how often the board meets but also hoe the UNECE seems to be applying fixed, metric based testing criteria to even hands on ASAS systems. For example, it appears from the latest draft of the regulations that when approaching a pedestrian at speed, the vehicle is not permitted to leave it’s lane to avoid a collision, even if this is the only way to avoid a collision.
Surely tests should be defined at a more high level? Ie “At 70Mph does the ADAS system safely avoid pedestrians spaced x meters away” Putting restrictions on the how (do not change lanes etc) will restrict the number of accidents and/or fatalities that such systems could have otherwise avoided.
The other concern is that since Tesla’s new FSD v12 software is trained entirely with video and no parameters as such, how would Tesla adopt the system to rules which are very statically defined and not necessarily in-line with how real-world drivers use their vehicles?
Wrap Up
Elon has stated he wants to prioritize International markets for FSD in the next 6 months and the UNECE regulations appear to be quickly perusing an early 2024 deadline for hand-on ADAS systems.
So it feels more likely that FSD with steering wheel nags will be available to the UK markets in early 2024.
There are technological, regulatory and possibly political issues that could get in the way of this date but it’s a significantly better position than being left in the unknown for so long!
If anyone has more information on the UNECE process or the seemingly strange restrictions in their draft proposals and why they are worded in such a way, please do get in touch, we’d love to have you on the podcast! studio@evshowpodcast.com
