zeoccl.com

China's Autonomous Driving Era: Tech, Vision, and the "True Path

Five years ago, the capital market's trend-setters, the enthusiasm of the competition between new forces and established manufacturers was fervent. Five years later, the end-to-end industry is thriving and passionate, with autonomous driving and cockpits accelerating restructuring, and the enthusiasm remains.

Looking back at the past and looking forward to the future, this is a grand industrial story that belongs to China's autonomous driving industry.

In April 2012, from Silicon Valley to Beijing.

Yu Kai, who had made significant achievements in the field of machine learning, resolutely left the NEC laboratory in the United States, leaving the laboratory he was familiar with, and stepped from the forefront of artificial intelligence theory to the front line of AI applications.

It is not a new thing for scientists to enter the industrial sector. Two years later, at the University of Missouri, Han Xu, who had already obtained a tenured professorship, also made the decision to leave academia and take root in the industrial sector.

Advertisement

The integration of talent and capital has made Baidu the most suitable hotbed for the gestation of autonomous driving. In addition to the well-known Yu Kai and Han Xu from the academic world, there are also two senior engineers from Google - Peng Jun and Lou Tiancheng.

In 2013, Baidu's autonomous driving car project officially started, gathering the top software engineers in China, and steadily promoting the rapid advancement of China's autonomous driving industry over the past decade.

Perhaps even Baidu itself did not expect that, in addition to building the Apollo platform to serve multiple clients within the autonomous driving industry, as the Whampoa Military Academy of Chinese self-driving, it has also sown the seeds of hope for the autonomous driving track.

In 2015, Yu Kai left Baidu, and Horizon Robotics was born, becoming China's first company to propose independent research and development of artificial intelligence chips and the first to achieve mass production of wafers.

"To truly achieve the popularization of artificial intelligence and accelerate its efficiency, relying solely on software is not enough. We should be more aggressive and design specialized chips for artificial intelligence."Yu Kai described his original intention when he talked about the establishment of Horizon. After three years of deep cultivation in the field of autonomous driving, scientists have further understood the theory and application.

Practitioners like Yu Kai have chosen to accelerate into the entrepreneurial track - in 2016, Peng Jun and the well-known Lou Tiancheng in the industry co-founded Pony.ai, and in 2017, Han Xu founded WeRide, making the autonomous driving track blooming with various flowers.

Now, all roads lead to the same destination. After eight years of diverging lines, Horizon, Pony.ai, and WeRide have all stood on the IPO stage.

The bell is about to ring, and behind the交错 of cups, this is also the history of the transformation of China's autonomous driving industry.

1. In 2024, autonomous driving "finally bears fruit"

The story starts from the current time point. In the first half of 2024, the proportion of new car deliveries equipped with ADAS exceeded 60%.

This means that autonomous driving has completed the penetration of the automotive industry and is no longer just a theoretical discussion for the public. Assisted intelligent driving has exploded in the past five years, becoming an indispensable dimension in the evaluation of new cars, and the price has gradually become more affordable in the "involution" of China's automotive industry.

Pull the time back to a year ago, from July 2023, WeRide, RoboTaxi, and Pony.ai have successively appeared in first-tier cities such as Guangzhou, Shanghai, and Wuhan, and unmanned taxi has also entered reality from science fiction stories.

In the eyes of the market, autonomous driving has finally completed a heavy node leap - after the product verification from 0 to 1, it moves towards the expansion and transformation from 1 to 10.

The day and night efforts of the top players are the "rocket head" of the rapid development of autonomous driving in the past five years, but the hardships involved are not enough to be told to outsiders. Among them, the automotive-grade AI chip is the "first pass" that Chinese autonomous driving companies have encountered.From the perspective of training requirements, the development and testing of autonomous driving technology are a bottomless pit in terms of computing power. The road environment is complex, and the fusion of multiple sensors leads to an exponential increase in the demand for computing power in deep learning. For a product to truly land, it requires tens of thousands of training and verification cycles. The demand for computing power for L3 level autonomous driving is 20-30 TOPS, L4 level requires more than 200 TOPS, and L5 level exceeds 2000 TOPS. This无疑是 a huge challenge for chip specifications at the time.

Horizon was born for this issue and has been struggling for a long time in front of this 'tough bone'.

"The second half of 2015 happened to be a small winter for financing, and Horizon obtained a good financing. At that time, I joked with my friends about how to spend this money? Later, after our chip team started, we found that this money was far from enough." Yu Kai once described the situation to the media when recalling the early days of making AI chips. Yu Kai may not have expected that every time a car-grade AI chip is taped, the R&D investment will exceed 50 million US dollars.

In the early days of Horizon, the company was still wavering between the automotive and Internet of Things industries, and it took more than a year to focus on the automotive industry, making automotive AI chips the 'only target' for Horizon. Under such economic pressure and internal turmoil, Horizon's first car-grade AI chip was released in 2019 after several rounds of financing.

The Journey 2, crowned as China's first AI chip, is equipped with the independently developed high-performance computing architecture BPU2.0, providing 4 TOPS of equivalent computing power, with a typical power consumption of only 2W. Even though there is still a gap compared to Nvidia's Orin launched at the same time, it also gives domestic new car forces a hope to build ADAS.

However, Tesla's FSD chip has given Horizon some acceleration invisibly.

When Tesla's large screen is equipped on the narrow space of civilian cars, people may think of the sentence Jobs said when releasing the iPhone 4s, 'Why do we need a keyboard?' With the example of Nokia being slapped on the beach by Apple, the top car companies have no time to explore the rationality of the intelligent system. Not falling behind has become the biggest goal for domestic new car forces. Even as a rising star, as a local supplier, Horizon is passively 'betting'.

The cooperation between Ideal and Horizon in 2020 attracted attention. At this time, Xiaopeng and NIO chose to join hands with Nvidia's Orin chip, while Ideal bet on Horizon. From the timeline at the time and the operating conditions of the two companies, this is almost a gamble - the combination of 'new brand + new chip', neither Ideal nor Horizon can afford to lose.

Looking back at this bet, the winner is obvious. The performance of the Journey series on Ideal models has certified that the computing power of domestic AI chips in intelligent assisted driving is not inferior to the top companies. Subsequent market actions have also verified this fact, and Xiaopeng and NIO have subsequently invested a lot of resources into the path of chip self-research.

On July 27, 2024, NIO's "Shen Yu NX9031" chip was released, with 50 billion transistors, claiming to be the world's first 5nm intelligent driving chip, becoming NIO's answer to the autonomous driving scenario. A month later, Xiaopeng quickly followed the team. After four years of dormancy, the "Xiaopeng Turing" chip was successfully taped, covering scenarios such as AI cars, AI robots, and flying cars. This chip, facing the L4 self-driving scenario, is equipped with a 40-core processor, 2 independent image ISPs for vehicle perception and user-perceptible images, and 2 NPU for processing neural network data, capable of running 30B large model data.Ideal aims to "grab with both hands," holding hands with the horizon while not giving up on the self-development process of AI chips. Currently, the Ideal L9 Pro is equipped with the Journey 5 chip, which can simultaneously support the industry's autonomous driving algorithms (BEV) and high-speed NOA. On the other hand, Ideal has also invested a significant amount of resources in the independent development of intelligent driving SoC chips. According to insiders, this chip, named "Shu Ma Ke," has poured Ideal's in-depth research into Chiplet and RISC-V technologies, and is expected to be released by the end of this year.

The breakthroughs of "Wei Xiao Li" in AI chips are a microcosm of the leapfrogging of China's new car-making forces in autonomous driving. It can be said that on the most fundamental "neck-grabbing" issues, Chinese self-driving is emerging from the predicament.

In addition to AI chips, from sensors to perception algorithms, from high-precision map services to cloud map manufacturers, the chain of the autonomous driving industry is gradually being unblocked, and costs are gradually being reduced, paving the way for autonomous driving services to reach the public.

Upstream and downstream enterprises are all feeling their way through the river together, and the backbone forces have completed the great escape in the continuous process of burning money, in order to achieve the high song and strong progress of the autonomous driving industry in the past five years, and stand on the stage of the 2024 IPO.

Ten years ago, mobile phones replaced PCs and became smart devices that everyone could reach. And in 2024, it also stepped down from the altar, no longer mysterious, and became an accessible emerging technology. Now, there is a new protagonist on the stage - autonomous driving.

II. 2000 days: visible forks in the road, hidden intersections

"Achieving autonomous driving is like climbing Mount Everest. At least there are two routes available for us to choose from, the south slope and the north slope, and there will always be pioneers and followers. Although there is only one summit, the experiences of climbers and the stories during the climbing process are often endless, fascinating, and fascinating."

This is the current state of autonomous driving described by Zhang Yaqin, an academician of the Chinese Academy of Engineering and Dean of the Institute of Intelligence at Tsinghua University, in a speech in October 2023.

The difficulties of climbing Mount Everest cannot hold back the brave, but instead inspire the inspiration of various solutions for autonomous driving players.

It can be said that in the past five years, the smoke of lidar and visual algorithms has never stopped, and the safety and commercial feasibility of the two strategies have been continuously hammered.Laser radar solutions with a lower difficulty coefficient transform the autonomous driving problem into a cost issue. By combining high-precision maps with real-time distance and speed measurements of the scene, these solutions evaluate road conditions for autonomous driving through the integration of prior information and real-time data.

In such solutions, the system is equipped with high-precision maps of cities, recording the shapes and distances of static buildings. Real-time 3D point cloud data are generated by lidar or millimeter-wave radars, capturing current dynamic data. Synchronization of localization and mapping (SLAM) technology is then relied upon to establish environmental maps, providing vehicles with planning and decision-making information. Since high-precision maps already provide most of the prior information, this solution is considered more reliable and safe, becoming the preferred choice in the autonomous driving industry. Companies like Waymo in California, USA, and Luobo Kuaibao in Pudong, Shanghai, have adopted this solution.

As domestic prices have "rolled" down, the implementation of lidar solutions has also taken off like a rocket.

Leading companies such as Gaode Map, Tencent Map, and Baidu Map, as the top five high-precision map providers, hold over 80% of the high-precision map market, covering multiple leading automotive companies. With the domestic unmanned driving market surpassing the 10 billion yuan mark, the cost of high-precision maps has gradually decreased, with leading manufacturers quoting in the hundreds of yuan per vehicle, reducing the extent of map collection, drawing, and maintenance year by year.

The price war has brought dividends to lidar solutions, but such industrial upgrades are not without challenges. In 2021, accidents involving multiple brands of cars occurred successively, leading to a decline in public trust in the safety of lidar-based autonomous driving.

On August 12, 2021, a NIO ES8 car was involved in a traffic accident on the Shenhai Expressway after activating the NOP pilot assistance feature. On the same day, a Xiaopeng G3 rear-ended a stationary vehicle at a speed of 70 km/h with the ACC function enabled.

The occurrence of black swan events has also led to a cold winter in autonomous driving financing. In 2022, there were 125 autonomous driving investment events involving more than 20.5 billion yuan, although the total number of events was on par with 2021, the cumulative disclosed amount was less than one-third of 2021. As safety accidents continue to emerge, the capital market is wary of the future of autonomous driving.

However, safety is only the surface of the sugar-coated shell of autonomous driving; the lack of core profitability has also caused investment institutions to pause their spending.

Overall, the industrialization of products with lidar at the core is advancing faster than those based on visual algorithms, but the goal of commercialization is still far from being achieved. Taking Waymo, which was the first to operate Robotaxi in California, USA, as an example, its Q2 2024 revenue was $365 million, higher than the $285 million of the same period last year, while losses expanded from $813 million last year to $1.13 billion. This means that even if companies in the lidar industry continue to reduce costs, "burning money" will still be the mainstream in the future.

Under this background, autonomous driving centered on visual algorithms has become a new stage for major companies to compete.Visual algorithms, compared to LiDAR, are more like a game for the "wealthy." To break free from the prior information of high-precision maps and rely on cameras and sensors to perceive everything around, is an extremely challenging task for visual algorithms. Choosing this path requires a decade or even longer investment of funds, recruiting top-notch talent and equipment, with a technical difficulty level higher than that of LiDAR.

It is clear that under such financial pressure, only large enterprises with substantial funding can remain at the table in this race. Baidu and Huawei, with their deep foundations in domestic autonomous driving, and Tesla overseas, are all leading players who have been deeply involved in this race.

In the long term, if visual algorithms can truly be delivered, their maintenance costs are inherently lower than products centered around LiDAR, making them more likely to achieve commercialization quickly.

However, Tesla's stock price drop of 9% on October 10, 2024, demonstrated the capital's "disapproval" of Tesla's answers. Essentially, capital requires a complete and clear business model, detailed and rich technical details, which are values that cannot be conveyed by throwing a party. Such demands are not only directed at players like Tesla who are deeply involved in visual algorithms but also at every player in the autonomous driving race.

As Academician Zhang Yiqun described, autonomous driving players from different routes are heading towards the same mountain peak in 2024.

This mountain peak is an autonomous driving product that can withstand scrutiny, and a commercialization model that can be digested.

Ten years of high capital investment, five years of ultra-high-speed development, whoever can climb this mountain first, turning continuous "burning money" into "monetization," will be crowned as the king.

III. End-to-end long-termism, and the "AI Agent" node of autonomous driving

Twenty years ago, the steady progress of the autonomous driving industry was like a marathon. Now, five years of rapid growth do not represent the final sprint.

The emergence of GPT has given artificial intelligence a "substance," a "direction" for true natural language processing, and has also made large models a productive tool that can be applied to various industries."End-to-end" is quietly becoming the new consensus in the market for a new paradigm in autonomous driving.

The past modular deployment focused on four directions: perception, prediction, planning, and control. For different manufacturers in the industry, it was only necessary to optimize in one direction, or even a sub-direction, and link with other modules' technical solutions through protocols and interfaces to obtain a mature and deployable autonomous driving solution. This industrial upgrade route was easy to implement and also improved the development efficiency of technological iteration.

However, the shortcomings are also obvious. The limitations of modularization were not taken seriously in the early stages of autonomous driving. Although sub-tasks were accompanied by information loss, resulting in local optimal solutions, leading to cumulative errors, such systemic issues could not be resolved within the existing architecture.

The transformation of the AI Agent is a turning point.

AI researchers began to realize that the increase in data volume and computing power could cause a qualitative change in the system, and the power of large models exceeded expectations. At the same time, in the industrial transformation of autonomous driving, the integration of modules also constructed some small "end-to-end" tasks, which obviously performed better than purely modular solutions. For example, replacing the perception module with a BEV combined with a Transformer solution introduced more data but reduced engineering effort, and also achieved better performance.

What is the charm of "end-to-end"?

To achieve Level 4 autonomous driving, it is unrealistic to reason and process all possible scenarios. However, if perception-prediction-decision-making-control is considered as a complete module, using sensors to obtain raw data and directly outputting vehicle driving actions, then the so-called step-by-step reasoning is not necessary. Based on a large number of data samples, the entire neural network will be driven by data, with operators calculating various situations, and it also has better generalization without data loss.

Concretely, tens of thousands of lines of code in the past, with the support of large models, now only require a few hundred lines but achieve better results.

Such a product architecture has gained consensus in the autonomous driving industry and has become a template for many AI products. For example, industries such as robotics and human-computer interaction currently recognize the end-to-end processing model, which is the most popular technical direction at present.

Building such an autonomous driving solution also allows autonomous driving companies to no longer be troubled by the form of delivery. Whether it is a Robotaxi or an autonomous driving solution for civilian vehicles, under the "end-to-end" technical model, as long as enough real data is collected, the entire chain can be run through, achieving delivery in various scenarios.This is also the core step of domestic automobile manufacturers this year—Xiaopeng released end-to-end large models for autonomous driving in 2024—XNet perception neural network, XPlanner planning control large model, and XBrain large prophecy model. They are respectively in charge of perception, planning, and decision-making, working together to handle complex scenarios within autonomous driving. In addition to Xiaopeng, Huawei, SenseTime, and DeepRoute also respectively handed in their answers for end-to-end solutions to autonomous driving.

However, looking back at the current end-to-end solutions, they still have the shadow of modularization—Xiaopeng's three-party collaborative work method did not achieve the original intention of end-to-end integration, and SenseTime's theoretical model lacks the support of real data. Such solutions are closer to domestic manufacturers' attitudes towards end-to-end applications and still have a gap from services that can be implemented.

On the other hand, the biggest pain point of end-to-end is the amount of data. Essentially, large end-to-end models are a trade-off between data and computing power and engineering difficulty. If computing power can be solved with funds, then the amount of data needs to be accumulated in reality. Building a large model for various autonomous driving scenarios requires an excessively large amount of data, and if enough data is accumulated, it will become the next proposition for autonomous driving manufacturers to explore end-to-end.

Looking back at the growth of the entire autonomous driving industry, after the intersection, a mature commercial monetization model still needs to be combined with long-termism, showing the universality of autonomous driving solutions, and achieving continuous progress in autonomous driving technology.

In 2023, HexaTech and Black Sesame Intelligence, which are in the upper and middle reaches of autonomous driving, went public one after another. Now, the baton of going public has come to the hands of Horizon, Pony.ai, and WeRide in the middle and lower reaches of the industry, marking another important turning point in the autonomous driving industry.

Five years ago, the trend setters in the capital market, the competition between new forces and old manufacturers, were full of enthusiasm. Five years later, end-to-end is vigorous and warm, autonomous driving and cockpit are accelerating reconstruction, and the enthusiasm is still there.

Looking back at the past and looking forward to the future, this is a grand industry story unique to China's autonomous driving industry.

Share This Post:-