Tesla’s Full Self-Driving officially transitions out of Beta testing into Supervised era

Can pureblood FSD be far behind?

Author|Wang Lei


Editor|Qin Zhangyong

A detailed change brought Tesla's belief in autonomous driving to its climax.

With the 2024.3.10 version update, the word “Beta” after FSD has finally been removed.

Beta version, literally translated as test version or experimental version, is mainly used to collect user feedback, discover and fix problems before the product is officially released.


In 2021, Tesla officially released the FSD Beta version. Since its release, this naming method has been controversial. Many people attacked the point that the naming method is not rigorous. Now Tesla cannot achieve fully autonomous driving, and There is also suspicion of using car owners as guinea pigs.

Now that the FSD test program is over, Tesla has turned to a very “obscure” word to replace Beta——Supervised.

Supervised means supervision. Compared with FSD Beta, FSD Supervised is more like a rebranding process, which means Tesla is more confident in its autonomous driving capabilities.

And what’s interesting is that in the past ten days, FSD has experienced an unprecedented burst of updates, with five updates in eight days.

According to officials, under the supervision of the car owner,The latest version of FSD Supervised lets you drive a Tesla almost anywhere:

It will perform lane changes, take turnoffs to follow your navigation route, navigate around other vehicles and objects, and turn left and right.

Of course, Tesla remains as cautious as ever. It said that even if FSD Supervised is turned on, you should be extra careful and stay focused because “it will not make your vehicle fully autonomous.”

In any case, FSD has been de-Betaized, and the real FSD is coming soon.

01,Small changes, big breakthroughs

From the beta version to the regulatory version, although it is just a simple name adjustment, behind it is a deep innovation in a substantial sense.

As early as July 2016, in the “Second Chapter of Tesla's Ambition”, Musk said when the suffix of the Beta version would be removed:

It's important that we explain why Autopilot is called 'Beta'. It is not used here in its usual sense when describing 'beta software'. Each version undergoes extensive internal testing before being released to consumers. It's called a Beta to reduce blind trust among consumers and to show that it will continue to be improved. (Autopilot is always off by default)

Next comes the point.

“Once Autopilot reaches the point where it is safer than the average American carabout 10 times(Musk believes), the Beta test version logo will be cancelled. “

According to the latest safety data disclosed by Tesla, with the FSD function turned on, Tesla vehicles are only likely to have an accident every 5.39 million miles driven, which is far lower than the national average of one accident every 670,000 miles driven.

In self-driving mode, Tesla’s safety factor is equivalent to that of the average American driver 8.04 timeseven in basic driving conditions, the safety level of Tesla vehicles is approximately 1.49 times higher.

Although it has not yet reached the “10 times safer” goal proposed in the second part of Tesla's master plan. But obviously, the fact that FSD has canceled the Beta suffix and opened it to FSD users has shown that Tesla internally believes that the FSD system is relatively mature and can be delivered to car owners.

In order to pave the way for this update, and perhaps to expand the penetration of FSD, Musk gave Tesla users a “head-click experience.”

As early as a week ago, Musk stated on function.

However, in order to prevent abuse of the auxiliary system, Tesla has specially produced an auxiliary driving instruction video. Before using the FSD fully automatic driving function, car owners must watch the auxiliary driving instruction instructions before using it.

Not only that, Musk also sent a letter to all employees, requiring Tesla sales in North America to take users on short-distance test drives of FSD before delivery, even if it affects the speed of new car delivery.

Musk stated in the letter that in the future, in North America, it will be necessary to install and activate Fully Self-Driving (FSD) V12.3.1 and allow customers to conduct short-distance test rides before delivery.Mandatoryof. Few people really realize how well (supervised) FSD actually works. “I know this will indeed affect the delivery process, but we still need to ensure that the intelligent driving experience is mandatory during the test drive.”

Immediately afterwards, Musk sent a second letter to all employees, requiring Tesla service centers to provide users with FSD test drive experiences after completion of repairs and services.

Later, Tesla chief designer Franz also tweeted that the delivery center team must ensure that everyone can get the new car and FSD 12.3 test drive experience.

The reason why the attitude is so tough is that the penetration rate of the “perfect FSD” that Musk believes is too low.

According to statistics from the research organization Troy Teslike, starting in 2019, Tesla made a promise to achieve autonomous driving. This year, the popularity of FSD reached its peak. 46%.

But five years have passed, and the penetration rate of FSD, which has experienced multiple delays, is getting lower year by year. In the second quarter of 2021, the penetration rate of FSD was only 11.1%. By the third quarter of 2022, the global penetration rate of FSD was only 11.1%. 7.4%of which the penetration rate in North America is 14.3%, the penetration rate in Europe is 8.8%, and the penetration rate in the Asia-Pacific region is even lower, only 0.4%.

This update to cancel the test suffix happens to be within the one-month trial period. Musk's calculations can be said to be quite loud.

02,Can true autonomous driving be far behind?

“I think Tesla will have a ChatGPT moment, if not this year, I think it will be no later than next year.”

Musk's confidence in saying this comes from Tesla's major breakthrough in autonomous driving technology. FSD V12 has end-to-end AI, which can input images and then output steering, acceleration, and brake light actions.

And in this update note, “FSD (Supervised)” is placed in the most conspicuous position of the 2024.3.10 version update note, and the concept of FSD V12 is officially proposed.

Compared with the previous version, the biggest improvement is the FSD (Supervised) V12 updateUpgrading the urban road driving stack into a single end-to-end neural networkthis neural network was trained on millions of video clips and replaced more than 300,000 lines of C++ code.

For Tesla FSD, the V12 version is a critical node.

As early as mid-January this year, Tesla pushed the end-to-end FSD version number V12.1.2 Beta to some car owners. After the version was pushed, overseas Tesla owners and video bloggers uploaded some test videos and evaluations. The video itself is mediocre, but what really deserves attention is the “end-to-end” aspect.

Before FSD V12, Tesla’s autonomous driving system had always relied onRule judgment, any behavior has rules and conditional judgment codes. When the car's camera recognizes lanes, pedestrians, vehicles, signs and traffic lights, it then uses hundreds of thousands of lines of C++ code manually written by Tesla engineers to deal with various situations, such as stopping at a red light, going on a green light, and running when there is not enough traffic. Pass through intersections only when fast vehicles collide, etc.

It’s not just Tesla. In traditional autonomous driving models, rules account for a relatively high proportion. If you want to improve model performance, you need a large number of excellent engineers to write a huge amount of rule code. The upper limit of the model also depends on the quality of the rule code.

The FSD V12 version has ushered in a new stage at the algorithm level.Abandoning the previous methods that relied on manually coded rules and machine learning models, we instead fully adopted an end-to-end neural network AI system.. For comparison, FSD V12 has only 3,000 lines of C++ code, and FSD V11 has more than 300,000 lines.

Tesla FSD's end-to-end solution essentially compresses tens or even hundreds of millions of video clips, including human driving habits, into the parameters of an end-to-end neural network.

The neural network is like a human brain. It can think on its own in an AI way, directly input sensor data when encountering a scene, and output steering, braking and acceleration signals, and there is no coding in this process.

This new architecture enables the software system to learn to drive like a human driver, continuously optimizing and improving its own driving decision-making capabilities by observing and imitating the reactions of excellent drivers in various complex situations.

All humans need to do is give the neural network“Feeding” videoallowing it to continuously learn, optimize parameters, and “learn” to drive by analyzing billions of frames of video of how humans drive cars.

As early as Tesla's second quarter earnings conference call in July 2023, Musk once introduced the end-to-end FSD training method. It took Tesla about a quarter to complete the training of 10 million video clips. . “Training 1 million video cases, it can barely work; 2 million, it's slightly better; 3 million, it will feel Wow; when it reaches 10 million, its performance becomes incredible.”

Of course, there are definitely more than 10 million training videos, but in this way,The more you feed, the more anthropomorphic your FSD becomes..

Although the development of smart driving in China can be said to be blooming everywhere, it has to be admitted that Tesla is still ahead of all car companies.

The current mainstream intelligent driving systems almost all adopt modular models, that is,Perception, prediction, and planning are divided into three independent modelseach model is relatively independent, and the downstream planning model requires engineers to write a lot of code to formulate driving rules. This is why the current smart driving does not feel “anthropomorphic”.

But good food never comes late. Since Tesla announced the FSD V12 version, the “end-to-end” craze has also swept the domestic smart driving circle. At the same time, network occupation by BEV, Transformer, and Occupancy have also become hot words in the industry.

At present, in China, including Huawei, Wei Xiaoli, Baidu and other car companies or Tier 1 companies, they all want to make efforts in end-to-end, but it is not yet known when they will be fully implemented. After all, Tesla’s push in the United States is also It’s still a “regulated” functional experience.

Similarly, the end-to-end solution for autonomous driving is also the key to FSD's ability to enter China. Previously, data collection and model training were the main factors that prevented FSD from entering China.

The end-to-end solution undoubtedly breaks this limitation and means that Tesla does not need to build an intelligent driving algorithm team of thousands of people in China. Only an operation team consisting of a local operation team of about 20 people and a data annotation team of hundreds of people can ensure the implementation and operation of FSD in the country.

So, when FSD really enters China, will it trigger a new smart driving storm?