"Understanding the 'Magic' of Self-Driving Software" Everything you need to know about E2E self-driving (FSD V12) in one article!
This article uses auto-translation in part.
Last week, Musk commented on X (formerly Twitter) that FSD is progressing well. He mentioned that 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 has progressed to the point where it should be called V13. Furthermore, he stated that the other two upcoming releases in the pipeline should be called V14 and V15, respectively. He added that if known bugs are fixed, it will take over a year of driving for FSD to trigger manual intervention. $Tesla (TSLA.US)$He mentioned that when the known bugs are fixed, it will take over a year of driving for FSD to trigger manual intervention.
Tesla in the US is completely overhauling its self-driving software with the release of FSD V12. The keyword is "End-to-End (E2E)". Since the release of FSD V12, the speed of FSD updates has been increasing rapidly, making significant progress each time.
Such technology is expected to "unlimitedly improve accuracy, dramatically reduce the cost of self-driving, and if realized, it will be difficult for existing automotive giants to compete with Tesla on self-driving software."
What is E2E?
If you want to know specifically how E2E differs from previous self-driving software, first you need to understand how self-driving software has operated so far.
Firstly, structurallySo far, the previous self-driving systems have taken a submodule approach. The AD system is divided into perception, planning, and control, first accurately perceiving the surrounding dynamic and static traffic participants and road network structure, then planning the vehicle's travel path, and finally, controlling the vehicle through actuators in a closed-loop manner. In such a system, clear interfaces and interfaces are designed between modular parts that model human cognitive steps.
On the other hand, in Tesla's FSD V12 E2E, rather than such a mechanism, the disconnection between modules like perception, planning, and control is eliminated, and key modules are combined to form a large neural network.
Secondly, formallySoftware using the submodule approach takes the form of a combination of manual coding and neural networks, with a high proportion of manual coding. Especially in the fields of regulation and control, most automobile companies still rely on rule-based traditional algorithms and manual coding.
On the other hand, Tesla's FSD E2E solution is implemented using a full-stack neural network that directly inputs sensor data and outputs steering, brake, and acceleration signals. In theory, the entire process can be achieved without any coding.
The third, in principleEnd-to-End Large Model extracts driving knowledge by compressing a vast 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 like GPT compresses internet-level data. Similar to human experience, driving knowledge is distilled and imprinted on the brain's neurons and synapses through various experiences in life.
Finally, in terms of development approachWhen considering, the FSD V12 of the full-stack neural network is a product of the Software 2.0 era, completely data-driven. Once the neural network's layers, structure, weights, parameters, activation functions, and loss functions are fixed, the quality and size of the training data become the sole factor determining the performance of the end-to-end neural network.
On the other hand, submodular systems are positioned between Software 1.0 and 2.0. Apart from using neural networks in certain parts, other parts utilizing manual coding still depend on the benefits of design rules and the performance of conventional algorithms.
Advantages and disadvantages of E2E
Generally, E2E is believed to have the following main advantages: ① High technical ceiling, ② Data-driven solutions for complex long-tail fit issues, and ③ Elimination of serious modular cumulative errors.
On the other hand, like generative AI such as the recently popular GPT, this technology also has features such as 'lack of interpretability' for those created by its internal operations, and high barriers to participation due to the need for large amounts of high-quality data and immense computing power.
Source: moomoo, Bloomberg
moomoo News of Zeber
This article uses auto-translation in part.
moomoo News of Zeber
This article uses auto-translation in part.
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