After 7 years, Tesla autonomous driving wrote a cool counterattack article

A major Easter egg for Tesla’s autonomous driving appears.

Foreign hacker master GreenTheOnly discovered a hidden code in the code of Tesla FSD Beta. It is understood that this code can open a new mode "Elon Mode" (Elon Mode, the code is named after Musk).

If the vehicle speed is below 60km/h and this mode is turned on, the vehicle can achieve L3 autonomous driving without the need for both hands to hold the steering wheel.

It can be said that Tesla is technically capable of achieving L3 autonomous driving. However, it has not yet received relevant approval from the government.

In the field of autonomous driving, Tesla leads the way.

However, leadership does not come overnight.

Reviewing the development history of Tesla's autonomous driving is actually no less than reading a "cool article" about counterattacks. Even Tesla has gone through five periods of self-transcendence, from lagging behind, to catching up and surpassing, to being crowned king.

In 2016, after Mobileye officially announced its "breakup", Tesla began to develop its own algorithms. For a period of time, its algorithm performance was not even as good as Mobileye's.

It was not until 2020, after reconstructing the algorithm, that Tesla established its position as the industry leader.

Technologies such as BEV, Transfomer, and automatic annotation that emerged during this period have now become technology routes commonly used in the industry today.

But Tesla did not stop there and upgraded the algorithm, such as adding timing information and upgrading to occupy the network.

From beginning to end, Tesla has not pursued the title of “first in martial arts.”

Tesla has only one goal: how to use algorithms to depict the real physical world in a purely visual solution to achieve autonomous driving.

In summary, Tesla adheres to the visual solution and starts from "first principles" and continues to iterate on algorithmic problems, making the algorithm more likely to understand the real world.

01. On the eve of the birth of the self-developed algorithm, Tesla and Mobileye broke up

On May 7, 2016, a car accident in the United States attracted global attention.

A Model S (2015), while using Autopilot, crashed into a white trailer coming vertically. The accident resulted in one death.

At that time, the accident was widely reported by the media as "the world's first fatal accident involving autonomous driving."

Under the overwhelming reports, Tesla's Autopilot has become the target of public criticism.

The cooperation between Tesla and Mobileye began in 2014.

In October 2014, Tesla released the first generation of hardware Hardware 1.0. Both hardware and software were provided by Mobileye, and the autopilot chip was Mobileye's EyeQ3.

Two months after the above-mentioned accident, in July 2016, Mobileye announced the termination of cooperation with Tesla. The supplier preemptively announced that the cooperation between the two parties had broken down, which is quite rare in the automotive market.

As for the reason for the breakup, both parties held different opinions at that time.

In Tesla's view, Mobileye's black box model was the reason for the breakup.

Tesla once bluntly stated in a document that under black box mode, Mobileye would have difficulty keeping up with the pace of Tesla product development.

Mobileye stated that it terminated the cooperation between the two parties because Tesla's autonomous driving function "exceeded the bottom line of safety."

In fact, the breakup between the two parties had already been foreshadowed.

In 2015, Tesla began to develop its own autonomous driving software and hardware, and it was only a matter of time before Mobileye was abandoned.

In April 2015, Tesla formed Vision, a software algorithm group based on computer vision perception, to prepare to develop its own software.

In the same year, Tesla also poached legendary chip designer Jim Keller from AMD. Subsequently, in 2016, Tesla began to form a chip research and development team, and Jim Keller served as the head of Autopilot.

Like many cliche love stories, Tesla also experienced a brief trough and loss after breaking up with Mobileye.

But in the days that followed, the frustrated Tesla eventually grew into a leader in the field of autonomous driving.

02. Tesla was in its infancy from 2016 to 2018

After bidding farewell to Mobileye, Tesla chose to develop its own full-stack autonomous driving algorithm to become independent and self-reliant.

In terms of the development of autonomous driving software and hardware, Musk has formulated a "hardware first, software update" approach for Tesla.

In terms of hardware, in October 2016, Tesla also released the second-generation hardware Hardware 2.0. The autonomous driving chip is provided by NVIDIA and is equipped with 8 cameras + 12 long-range ultrasonic radars + 1 front-facing millimeter wave radar, and this set of configurations continues to Hardware3.0.

In terms of algorithms, Tesla continues to use the conventional backbone network structure in the industry; uses 2D detectors for feature extraction; and manually annotates the data.

Overall, this set of autonomous driving algorithms is relatively primitive and traditional.

It is worth mentioning that during this period, Tesla’s autonomous driving algorithm was still in the technological catch-up stage.

From the perspective of hardware configuration, although HW2.0 is better than the HW1.0 previously provided by Mobileye, it was limited by the software algorithm. At that time, there was a big gap between Tesla's autonomous driving capabilities and Mobileye's.

Although Tesla launched HW2.0 in October 2016, after running idle for more than half a year, it was not until March 2017 that Model 3/Y began to be able to truly use the Autopilot function.

After its algorithmic capabilities caught up with Mobileye's, Tesla discovered that the currently used algorithms had many shortcomings. Among them, the most obvious is the issue of efficiency.

During that period, target detection for autonomous driving generally followed a common network structure:

Input → backbone → neck → head → Output

The backbone network is a feature extraction network, mainly used to identify multiple objects in images;

The neck is mainly responsible for extracting more detailed features;

After feature extraction, the detection head provides input feature map representation, such as detection objects, instance segmentation, etc.

It is worth mentioning that at that time, the autonomous driving visual neural network in the industry only had one head.

However, in autonomous driving scenarios, it is often necessary to complete multiple tasks simultaneously in a neural network, such as lane line detection, person detection and tracking, signal light detection, etc.

This makes the original algorithm have a situation where "the brain is not enough".

Therefore, in 2018, Tesla began its first innovation of the autonomous driving algorithm, targeting the autonomous driving network structure and efficiency.

03. From 2018 to 2019, the sharp edge of the algorithm was first developed.

In this algorithm innovation, Tesla built the multi-task learning neural network architecture HydraNet and used the feature extraction network BiFPN.

This improves the efficiency of Tesla's algorithm. Among them, the most distinctive one is HydraNet.

The word Hydra comes from the legendary creature "Hydra", so HydraNet is also called the "Hydra Network".

The reason for the name "Hydra" is that the HydraNet structure can complete multiple tasks instead of the previous single detection.

Compared with previous algorithms, HydraNet can reduce repeated convolution calculations, reduce the number of backbone network calculations, and can also decouple specific tasks from the backbone for independent fine-tuning.

However, this innovation is more of a "fine-tuning" of the algorithm and has not reached the level of reconstruction and leapfrogging.

In terms of fusion methods, Tesla still adopts a post-fusion strategy, the data is manually annotated, and the self-driving algorithm is still a small model. Compared with subsequent algorithm innovations, there is not much breakthrough.

During this period, after improving the traditional algorithm, Tesla also conducted a new round of hardware updates.

After four years of research and development, Tesla released the Hardware 3.0 system in April 2019. The biggest highlight is Tesla’s use of self-developed FSD chips.

Tesla's FSD chip has a computing power of 72TOPS, which is much higher than the self-driving chips on the market at the time. At the same time, the FSD chip is mainly composed of two NUDs, which has higher image processing efficiency and is not equipped with lidar.

The release of new hardware provides the possibility for the next iteration of Tesla’s algorithm.

After completing the preliminary work of hardware preparation, Tesla began an epic reconstruction of the autonomous driving algorithm.

04. Tesla’s autonomous driving will be the best in 2020

In August 2020, Musk tweeted that the Autopilot team was rewriting the underlying code of the software and reconstructing the deep neural network; a new training computer Dojo was being developed.

Musk’s tweet caused a stir. The market is paying attention to the development direction of Tesla's autonomous driving algorithm.

In his view, the rewriting of AP is not an optimization of the existing structure, but a "quantum leap."

Looking at Tesla’s self-developed algorithms in the past ten years, 2020 can be said to be its most brilliant year.

In this industry restructuring, a series of technical directions brought by Tesla have been used by the autonomous driving industry to this day, such as the combination of BEV+Transformer, feature-level fusion replacing post-fusion, data self-labeling replacing manual labeling, etc.

If the autonomous driving arena in 2020 is a period of competition among rivals, then since 2020, this arena has entered the era of Tesla.

(1) BEV+Transformer, autonomous driving enters the era of large models