Shulou(Shulou.com)11/24 Report--

The development details of Tesla's FSD V12 system have been exposed.

Although Musk predicted a change in the technical route of FSD V12, it was surprising that Tesla only began to train this neural network-based intelligent driving algorithm at the beginning of this year. Just four months later, the new system was ready to replace the old one, and eight months later, the new FSD V12 was unveiled live in Musk.

Behind this is a changed technical route, from rule-driven to data-driven, from modular design to end-to-end. At the same time, it brings new challenges.

FSD V12: only neural network generally speaking, Tesla FSD V12 has only one core feature: no rule code, only neural network.

What does it mean?

Most of the common autopilot systems in the market are designed in sub-modules, including perception, decision-making and control modules, and each task uses its own algorithm model. Among them, the AI algorithm is mainly used in the perception module, the decision-making and control module is still conventional, based on if else logic code.

That is, the code written by algorithm engineers will establish a set of rules for the autopilot system, such as stopping at red lights, passing at green lights, driving in the middle of the driveway, and so on.

So the disadvantage of this system is obvious, the rule setting standards are determined by the engineers, the driving style is easy to mismatch with the driver's habits, so the experience is very poor, it is better to drive by yourself.

Tesla FSD V12 only means that the previous perception, decision-making, control modules are not needed in the design, as long as determine the neural network architecture, and then input data for training.

A set of neural network can process all input signals and output driving decisions. Based on real human driving data, the system can learn how to drive and continue to drive better.

This is called from rule-driven to data-driven.

According to all kinds of environmental information input, the system judges how to drive in this case based on rules; when it comes to training, input human driving data first, and after the system fully learns human driving habits, judge how to drive according to the input environmental information in the actual driving environment.

If there is a situation that is not handled well, enter more data specifically for this scenario. This is similar to the ChatGPT training method, but is more suitable for the car version.

Before deciding to change the technical route, Tesla's autopilot team showed Musk that a neural network-based system could handle better situations.

When the road is littered with trash cans, fallen traffic cones, and some random obstacles, the car can accurately bypass the above obstacles, cross the driveway, and violate some traffic rules if necessary.

Before the live broadcast, Musk also tested FSD based on neural networks. For a total of 25 minutes, Musk stepped on the gas pedal only when the system was too cautious, but never touched the steering wheel, and once the system did better than he expected.

My human neural network failed here.

How to think about it in fact, this concept has sprung up among self-driving players long before Musk announced that FSD V12 became an end-to-end technology route.

Because the end-to-end autopilot system is easy to develop, you don't have to write a lot of code up front (there are more than 300,000 lines of C++ code in the FSD V11 version control stack), and you don't need engineers to design rules in advance. As long as you constantly enter human driving data, the system can watch and learn by itself.

But it also puts forward high requirements for autopilot players. For example, the input must be a large number of high-quality data, in order to better help the system to learn.

Musk found that when more than 1 million videos were entered, the neural network-based autopilot system began to perform well.

At the beginning of this year, Tesla already input 10 million human driving videos into this system, and they are still screened, the kind of veteran drivers.

Tesla's fleet of nearly 2 million vehicles around the world also provides about 160 billion videos for training every day. Tesla predicts that billions of videos will be used for training in the future.

This is a challenge for data volume, data annotation, computing, and so on.

And the reason why end-to-end technology is not on a large scale among self-driving players is that there is a key problem: it increases the unexplainability of autopilot systems.

At this stage, end-to-end autopilot is still a "black box", and there is no way to accurately explain why the system does not handle well in a certain situation.

So the solution given by Tesla is to find that the system is not handled well during the test, then feed more data pertinently.

For example, during Musk's live broadcast, the system almost ran a red light, and the solution is to enter more traffic lights, especially the video of the left turn signal.

In addition, Musk also set an indicator for the team to show the mileage of the FSD system in real time without human intervention. If there is an intervention, deal with the corresponding problem.

More importantly, learning in this way will give rise to a new problem: the system will learn not only the slippery operation of old drivers, but also the behavior of human drivers who do not comply with traffic rules.

For example, when you encounter a stop sign, more than 95% of people will pass slowly rather than stop completely. This means that regulators need to clarify regulatory standards.

The National Highway Safety Board is studying whether to allow autopilot systems that do not fully comply with traffic regulations.

In short, the introduction of Tesla FSD V12 is indeed of great significance to autopilot. Now that the whole process of AI can be realized, it is more likely to move towards AGI, that is, general artificial intelligence.

When will autopilot usher in ChatGPT? The gear of fate may start to turn from now on.

Reference link:

Https://www.cnbc.com/2023/09/09/ai-for-cars-walter-isaacson-biography-of-elon-musk-excerpt.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.