Compared with Tesla, large models become another "boost" for autonomous driving?

On August 3, Musk said on Twitter that Tesla is developing the "last piece of the puzzle" of FSD - "vehicle control" and is expected to achieve fully autonomous driving by the end of this year.

Musk is optimistic about the future of autonomous driving.

“Vehicle Control” is the last piece of the Tesla FSD AI puzzle.

Musk’s words indicate that Tesla will rely more on AI to control the vehicle rather than relying on hard-coded instructions to accelerate the implementation of Tesla FSD. The introduction of "vehicle control" will reduce more than 300,000 lines of C++ control code in the original FSD by about 2 orders of magnitude.

The timetable for fully autonomous driving is still controversial in the industry.

He Xiaopeng, founder of Xpeng Motors, said that high-end L4 or quasi-L5 may not appear until 2027 to 2030.

Tesla FSD is based on Transformer's BEV perception route and is the first time that large model technology has been applied to the autonomous driving industry.

Before ChatGPT became popular, large-scale model technology had already been used in "cars" and is now almost a must-have for smart driving.

Li Xiang, the founder of Li Auto, publicly declared that "the research and development and training of large models are necessary capabilities for smart electric vehicle companies."

There seems to be no doubt that the future of fully autonomous driving supported by large-model technology is coming. However, due to limitations in computing power, data, etc., it is difficult to reach a consensus on how long it will take to achieve this.

01 Can you "cross the river" by touching a large Tesla model?

Tesla's self-driving technology is entering a new stage.

Musk revealed that FSD V12 will no longer be a “Beta” when it is released.

Tesla allowed the outside world to see the potential of Transformer technology, and domestic car companies have successively followed this perception route.

Li Xiang once said that Li Auto’s investment in autonomous driving officially started after Tesla ran through the general logic.

Autonomous driving companies are following Tesla's lead in shifting to BEV architecture. The route that focuses on algorithms and uses cameras and various sensors as hardware has become popular. The main driving force behind this comes from the transition from high-speed NOA (Automatic Assisted Navigation Driving) to urban NOA.

At the same time, high-precision maps, once regarded as "hot cakes", are increasingly falling out of favor.

As early as 2019, Musk publicly stated: "Over-reliance on high-precision maps will make the autonomous driving system extremely fragile and make it more difficult to popularize."

This solution has higher requirements for chip computing power, sensor hardware, and algorithms, and car companies have gradually caught up with the technology accumulated over the years.

At the same time, from the perspective of technological advancement, omitting high-precision maps will help accelerate the popularization of urban NOA.

On the road of technological iteration, Chinese smart car companies are "crossing the river" by following the BEV large model road explored by Tesla.

The key problem with this seemingly easy path of imitation is that NOA, which is booming in China, has never left L2 assisted driving, while Tesla's FSD V12 claims to reach L4/L5.

Car companies are not shy about this.

According to Li Xiang's prediction, by the end of this year, the intelligent driving of most leading companies will be able to reach the level of Tesla at the end of 2021; by 2024, companies will generally be able to reach the level of Tesla in North America from the end of 2022 to the beginning of 2023. .

This means that the intelligence level gap between China's most technologically advanced car companies and Tesla is about 2 years.

Looking back at the process of obtaining Tesla's full autonomous driving capabilities, conquering Transformer technology is a key step.

In July 2021, Tesla FSD Beta demonstrated a new paradigm of autonomous driving perception based on BEV+Transformer.

The perception solution based on Transformer's BEV (Bird's Eye View) also allows Tesla to get rid of lidar and firmly follow the purely visual route.

The so-called BEV refers to using a neural network to map the image space to the BEV space, and converting the 2D images of the camera into a 3D scene to generate a bird's-eye view.

This picture with a "God's perspective" can realize 360-degree surround perception, and integrate time and space to form a 4D space.

BEV perception framework source: Essence Securities "Application of AI Large Model in Autonomous Driving"

Although two years late, Chinese car companies are accelerating to catch up with Tesla.

According to incomplete statistics, currently, car companies including "Wei Xiaoli", as well as Baidu Apollo, Huawei, DJI, Haomo Zhixing, Qingzhou Zhihang, Pony.ai, Yuanrong Qixing, Horizon, SenseTime and other autonomous driving companies Businesses are using BEV technology.

After trying the BEV effect of Transformer, Xpeng Motors began to build XNet.

Xpeng XNet can output 4D dynamic information (such as vehicle speed, motion prediction, etc.) and 3D static information (such as lane line position, etc.) under BEV.

Xiaopeng claims that this is the first mass-produced BEV sensing solution in China.

The latest urban NOA navigation assisted driving AD Max 3.0 released by Ideal will be launched in 100 cities before the end of the year.

This solution is also equipped with three neural network large model algorithms: static BEV network algorithm, dynamic BEV network algorithm and Occupancy network algorithm. This is a goal that many car companies are pursuing.

The most recent progress in the implementation of BEV technology comes from DJI.

On July 27, DJI Auto officially named its cost-effective intelligent driving solution "Chengxing". The "Chengxing Platform" realizes BEV perception with integrated driving and parking, and adopts OSP (Open Space Planning, open space decision-making and planning technology) , improve the traffic efficiency of intelligent driving in cities, highways, parking lots and other environments.

• 
  

• Previously, Baidu Apollo has upgraded visual perception to BEV perception, and UniBEV vehicle-side and road-side perception data are in the same coordinate system;

• 
  

• Huawei's ADS 1.0 is said to have implemented a Transformer-based BEV architecture, while ADS 2.0 further upgraded the GOD network, similar to Tesla's occupied network algorithm;

• 
  

• SenseTime has developed BEVFormer++, a surround-view perception algorithm for autonomous driving;

• 
  

• HaoMo Zhixing released HaoMo DriveGPT, the industry's first generative large-scale autonomous driving model, Xuehu·Hairuo, which established RLHF (Human Feedback Reinforcement Learning) technology by introducing driving data;

• 
  

• Qingzhou Zhihang's OmniNet realizes pre-fusion and BEV spatial feature fusion;

• 
  

• Pony.ai self-developed the "scalable network" multi-task large model BEV algorithm architecture, which can adjust the network size and its corresponding resource consumption rate based on different computing power platforms. Higher computing power can identify more static element types and dynamic obstacles. Subdivision of objects and a larger scope of identification;

• 
  

• Horizon released SuperDrive sensing fusion BEV technology;

• 
  

• NIO's NOP+ Beta version will be upgraded to the official version, using BEV and placeholder grid sensing models, and switching to the same NAD technology stack;

Although they may seem very different, the BEV technologies of each of the above companies are actually innovations based on the existing computing power, scenarios, and implementation speeds, with Tesla's BEV solution as the basis.

But BEV is just the first stop on the long march. Tesla has already achieved another technological peak: occupying the network.

At the end of 2022, Tesla will upgrade BEV to an occupancy network. The system's perception will change from 2D to 3D. It does not require secondary splicing and directly places the vehicle in the 3D world. The vehicle can receive the information within 10 milliseconds. The occupancy probability of each surrounding 3D location to predict the distribution of surrounding obstacles.

This greatly improves the vehicle's ability to cope with corner cases and improves generalization capabilities (that is, the AI's ability to recognize a class of objects after recognizing one object).

Comparison of BEV and occupied network effects Source: Essence Securities "Application of AI Large Models in Autonomous Driving"

The implementation of the NOA scene has made BEV popular, and the "removing high-precision map + occupying the Internet" model may become the next stop for car companies.

Although Tesla's FSD automatic driving system is rich in functions, it still requires the driver to take over, which means that FSD "has not yet reached the true L3 level."

From L2 autonomous driving to fully autonomous driving, the road ahead for large models of car companies is still long.

02 Towards the end of large model technology: aim at, become, and surpass "end-to-end"

In May this year, a paper published by the team of young scientist Li Hongyang of the Shanghai Artificial Intelligence Laboratory proposed for the first time a general model for autonomous driving that integrates perception and decision-making, and won the "CVPR 2023 Best Paper".

This is the first time in the 40-year history of the top conference CVPR that it has awarded the "Best Paper Award" in the field of autonomous driving.

Source: Shanghai Artificial Intelligence Laboratory

Li Hongyang's team proposed a goal-oriented autonomous driving algorithm solution (UniAD, Unified Autonomous Driving). Its design concept is to adopt an end-to-end architecture, take Planning as the ultimate goal, and integrate all autonomous driving modules.

Li Hongyang said that the difference between this solution and MTL, Tesla and other solutions is that the latter want to maximize the performance of all tasks, "while our solution only focuses on the results of Planning."