Find out what the “magical autonomous driving software” E2E autonomous driving (FSD V12) is in one article!
This article uses automatic translation for some of its parts
Last week, Musk commented on X (old Twitter) that FSD is progressing smoothly. He said FSD 12.4.1 has been released to Tesla employees. If all goes well, it is scheduled to be released to a handful of external customers by the end of this week.
Musk also commented that FSD 12.4.1 is advanced enough to be called V13. Furthermore, $Tesla(TSLA.US$It was stated that the other two releases in stages should be V14 and V15, respectively. Also, he added that if known bugs are fixed, it will take more than 1 year of operation for FSD to trigger manual intervention.
US Tesla will completely revamp autonomous driving software with the release of the FSD V12. The keyword is “end to end (end to end, E2E)”. Since the release of FSD V12, the speed of FSD updates has been getting faster and faster, and major progress has been made each time.
It is anticipated that such technology “has the potential to infinitely increase accuracy and dramatically lower the cost of autonomous driving, and if realized, it will be difficult for major existing automobiles to compete with Tesla on their own with autonomous driving software.”
What is E2E?
If you want to know specifically how E2E differs from automated driving software up until now, you first need to understand how autonomous driving software has operated until now.
First, structuralUntil now, autonomous driving systems have taken a submodular approach. AD systems are divided by perception, planning, and control; first, they accurately perceive surrounding dynamic and static traffic participants and road network structures, then plan the driving trajectory of cars, and finally, control the vehicle via an actuator in a closed loop system. In such a mechanism, clear interfaces and interfaces are designed between modules modelling human cognitive steps.
Meanwhile, in the Tesla FSD V12 E2E, there is no such mechanism, and disconnection between modules such as perception, planning, control, etc. is eliminated, and the main modules are combined to form a large neural network.
Second, formallySubmodular software takes the form of a combination of manual coding and neural networks, and the proportion of manual coding is high. Most car companies still rely on traditional rule-driven algorithms and manual coding, especially in the field of regulation and control.
Meanwhile, Tesla's FSD E2E solution is implemented using a full-stack neural network that directly inputs sensor data and outputs steering, braking, and acceleration signals. Theoretically, the entire process can be realized without coding.
Third, theoreticalThe End-to-End Large Model extracts driving knowledge by compressing a huge number of driving video clips. Tesla's FSD compresses human driving knowledge from tens of millions of video clips into neural network parameters, similar to how LLM such as GPT compresses internet-level data. Similar to human experience, driving knowledge is also distilled and engraved into brain neurons and synapses through various experiences in life.
Finally, the development approachIf you think about it, the FSD V12 full-stack neural network is a product of the software 2.0 era and is completely data-driven. After the number of layers, structure, weight, parameters, activation function, and loss function of the neural network are fixed, training data (quality and size) becomes the only factor that determines the performance of the end-to-end neural network.
Meanwhile, submodular systems are located halfway between software 1.0 and 2.0. With the exception of the part that uses neural networks, other parts that use manual coding still rely on the merits of design rules and the performance of conventional algorithms.
Advantages and Disadvantages of E2E
Generally, E2E is thought to have the main advantages of ① a high technical upper limit, ② a data-driven solution to complex fit long-tail problems, and ③ elimination of serious modular cumulative errors.
On the other hand, similar to generative AI such as GPT, which was a hit recently, this technology also has characteristics such as “lack of interpretability,” which is a black box even for those who created its internal operation, and high barriers to participation because large amounts of high-quality data and enormous computational power are required.
Source: moomoo, Bloomberg
— MooMoo News Zeber
This article uses automatic translation for some of its parts
— MooMoo News Zeber
This article uses automatic translation for some of its parts
Disclaimer: Moomoo Technologies Inc. is providing this content for information and educational use only.
Read more
Comment
Sign in to post a comment
Mr スコップ : Imif
183819413 : Like humans, E2E determines behavior according to rules of thumb based on perceptual information, and submodules determine behavior according to algorithms coded with sensory information similar to programming.
The former is the amount of data stored = comprehensiveness of rules of thumb = improved accuracy of action decisions, and compared to the latter amount of coding = comprehensiveness of algorithms = improved accuracy of action decisions, the learning speed is completely different. I don't know
HONDA N-ONE : AI can memorize traffic rules without permission and improve driving skills