news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Text | Titanium Capital Research Institute

Recently, Baidu's driverless online car-hailing service "Robot Run" has become a hot topic. Tesla is expected to release "Robotaxi" in August 2024. At the same time, domestic policies have been intensively introduced to stimulate the development of driverless driving. At the beginning of 2024, the Ministry of Industry and Information Technology and five other departments jointly issued a notice on the pilot application of "vehicle-road-cloud integration", covering mainstream first- and second-tier cities. Autonomous driving technology will be an important growth driver for the new energy vehicle industry. End-to-end architecture has gradually become the main theme of this year's competition in the field of intelligent driving. In order to compete for the leading advantage in the trend, automobile companies have begun to increase their research and development efforts.

What is the investment logic of the smart driving track? What is the development trend of end-to-end autonomous driving technology? Recently, Titanium Capital invited Dr. Liu Yudong, Executive General Manager of Chentao Capital, to share on related topics. He has long focused on investing in the smart driving track. He has worked at Geely Automobile Research Institute and Didi Chuxing, leading the mass production development of smart driving systems and forward-looking research on perception algorithms for a number of smart electric vehicles. He has rich practical experience in technology planning, product definition, team management, ecological cooperation and other fields. The host of this sharing is Wang Zeqing of Titanium Capital, focusing on new energy, new materials, AI, robot industry chain and transportation.

Chentao Capital’s investment logic in the smart driving track

1. Focus on intelligent driving

Chentao Capital is a private equity investment institution focusing on emerging industries, and smart driving is one of its key vertical tracks for investment. There are three main reasons for focusing on the smart driving track: first, as a field that combines transportation and artificial intelligence, smart driving has a huge market space, which is expected to exceed one trillion yuan; second, smart driving is a long-term track, and the technology and industry structure have not yet stabilized, providing start-ups with continuous entry opportunities; finally, the talent and technology scalability of the smart driving industry is good, and the flow of core technical talents brings investment opportunities in related industries such as AI and robots.

In the past seven or eight years, Chentao Capital has invested in nearly 20 smart driving related companies, most of which occurred in the seed round and angel round stages, and participated as a major shareholder. The investment areas are mainly divided into three sectors: the first is unmanned driving companies focusing on different sub-scenarios, such as mines, ports, urban distribution, sanitation cleaning, etc.; the second is the core supply chain related to smart driving, including wire-controlled chassis and upstream sensors and parts companies; the third is the software and service companies related to autonomous driving that have attracted much attention in the past two years, especially generative AI and data services.

2. The industry has started its second wave of growth cycle

Intelligent driving is seen as a long-term track, and China's autonomous driving market has experienced its first wave of growth since 2015. Although attention has declined in the past two years, this trend is in line with the Gartner curve of technological development, indicating that there is still huge growth potential and investment opportunities in the future.

At present, the intelligent driving industry is in an adjustment period after the first wave of investment peak and faces challenges such as commercialization, but the long-term prospects remain optimistic.We have a clear prediction of the development trend of technology and believe that the industry has entered the second wave of growth cycle. The signals of this cycle include the commercialization turning point of driverless companies and the listing of related companies, indicating that the intelligent driving industry has entered a new investment stage. Therefore, our fund started investing last year and continues to look for new opportunities.

Chentao Capital currently focuses on two investment directions: one is the commercialization of autonomous driving in different scenarios. We believe that autonomous driving technology will transform transportation and generate real value. The other is the cutting-edge technological breakthroughs and applications in the field of autonomous driving, especially end-to-end autonomous driving technology, which will be the main factor in the transformation of the autonomous driving industry in the next two to five years. We are highly sensitive and concerned about large models, 4D millimeter-wave radar, lidar technology, and various chip technologies in the field of autonomous driving. The following sharing will focus on different scenarios and end-to-end technologies of autonomous driving.

3. Investment logic for different scenarios of autonomous driving

We use the four-quadrant classification method to distinguish unmanned driving application scenarios based on speed (low speed, high speed) and load (carrying goods, carrying people). In terms of technical difficulty, high-speed autonomous driving is more difficult than low-speed, carrying people is more difficult than carrying goods, and open scenarios are more difficult than specific scenarios. Therefore, Chentao Capital initially focused its investment on the field of low-speed unmanned driving, involving market segments such as mining transportation, terminal logistics distribution, unmanned cleaning and port scenarios.

4. Commercialization progress of autonomous driving in different scenarios

We divide it into To G, TO B (enterprise market) and TO C (consumer market). In the TO G field, sanitation and security are the main application scenarios. Sanitation has achieved initial large-scale implementation, and security inspection robots are used in some cities.

On the TO B side, the commercialization of driverless vehicles in mines and ports has progressed the fastest, and a clear business model has been formed and has begun to be replicated on a large scale.

In addition, urban delivery scenarios have seen explosive growth this year, with leading companies announcing that they will expand their fleets to tens of thousands of vehicles, a more than tenfold increase in size. The rapid growth of urban delivery is due to finding suitable application scenarios in the delivery process, such as cooperation with express logistics companies and fresh food supermarket delivery. For B-side trunk logistics, since it is an open scenario, the technical difficulty is relatively high and it is still in the small-scale test operation stage.

On the C side, driverless applications are mainly Robotaxi (driverless taxi) and Robobus (driverless bus).

5. Driverless taxi: Progress summary of Carrot Run

Baidu's Carrot Run project is a positive signal for the disruptive effect of the entire autonomous driving industry.

There are three development stages of autonomous driving scenarios: unmanned, large-scale and commercialized.In the unmanned stage, Baidu has demonstrated its unmanned driving capabilities. Baidu has achieved the unmanned operation of RoboTaxi (driverless taxi) in 2022, conducted unmanned driving tests in Wuhan and other places, and supported the operation of driverless taxis through remote control driving centers. In terms of scale, Baidu announced the launch plan of RoboTaxi, and it is expected to launch about 1,000 vehicles by the end of the year. Some vehicles have been launched in Wuhan, showing initial signs of scale. The commercialization stage focuses on profitability. Liu Yudong believes that the profitability time of Robo Taxi is still unclear, which mainly depends on the safety ratio of remote control driving and the company's pricing ability in the future. At present, Robo Taxi provides ultra-low-priced services in the form of subsidies. Whether it can maintain the order scale while increasing the fare in the future is an issue that the industry needs to pay attention to.

The driverless industry is undergoing active commercialization attempts. Despite the challenges, technological progress and innovation in business models provide broad space for the industry's future development. The practices of companies such as Baidu not only promote the development of technology, but also provide valuable experience for the commercialization exploration of the entire industry.

The Carrot Run project has had a significant impact on the three levels of technology, industry and sector. Technical level: Carrot Run demonstrated a system solution that combines vehicle-side algorithms and remote control driving as a backup, providing a scalable solution. This model provides consumers with a true driverless experience through remote control driving when the vehicle-side technology is not yet perfect. In the future, more passenger car OEMs and other autonomous driving companies may refer to this solution to promote the widespread use of driverless taxis.

Industry level: The relatively large-scale application of Carrot Run provides a demonstration of industrial division of labor for the industry. It involves downstream operators and upstream component companies, bringing incremental benefits to these stakeholders. As operations mature, Baidu may further refine the operation links and asset holding plans to promote industrial division of labor, which is beneficial to the development of the entire industry.

Industry level: The breaking circle effect of Carrot Run has rekindled the public's interest in the autonomous driving industry. This event is the "ChatGPT moment" of the autonomous driving industry, which means that it is not only a technological breakthrough, but more importantly, it has attracted the attention of the public. With more people's attention, the autonomous driving industry will attract more resources, including talent and funds.

Development Trends of End-to-End Autonomous Driving Technology

The development of end-to-end technology has not only promoted in-depth research in academia and industry, but also brought new investment opportunities and research directions to the autonomous driving industry. Although end-to-end technology is still in the exploratory stage, its application prospects in the future autonomous driving field are broad and deserve close attention from the industry and investors.

1. End-to-end autonomous driving opens a new round of industrial revolution

End-to-end technology has received widespread attention from academia and industry in the past year, especially after the research results of the Shanghai Artificial Intelligence Laboratory won the CVPR2023 Best Paper Award, this technology has once again become the focus of academia. The end-to-end concept has been explored from 2016 to 2018. Tesla has made positive statements in this field. In August 2023, Elon Musk live-streamed FSD v12 based on the end-to-end architecture, officially starting to bring end-to-end to the market. In the past six months, China's leading smart driving companies have successively launched end-to-end autonomous driving plans. In May 2024, Wayve, a startup focusing on the research and development of end-to-end autonomous driving systems, received US$1 billion in financing. This is the first transaction with a financing amount of more than US$1 billion since the autonomous driving industry entered a trough in 2022, showing the capital market's confidence in this field.

2. Evolution of autonomous driving architecture

End-to-end is a technology that covers the entire process from sensor input to trajectory planning or control signal output, which is completely implemented by AI or neural networks.

The development of end-to-end technology can be divided into several stages: first, perception end-to-end, in which the perception module has been implemented by deep learning, but the decision planning is still based on rule definition; second, modularization of decision planning, trying to replace rule-defined decision planning with AI network; then modular end-to-end, what is transmitted between modules is the characteristics of image expression, and joint training and optimization can be carried out; finally, there is the end-to-end solution of a single neural network, which draws on generative AI and large model technology. Although it is less mature, it has great potential.

End-to-end is not the same as big models and world models, although end-to-end solutions may draw on the capabilities of big models or multimodal big models. Ideal Auto’s recently released end-to-end system uses a multimodal big model. World models are currently mainly used as a technical means of training data, and may become a key technology for autonomous driving in the future.

End-to-end technology is compatible with all types of sensors, not just pure vision systems. Although Tesla is the leader in pure vision systems, end-to-end solutions can also integrate different inputs such as lidar and millimeter-wave radar. Camera data is the easiest type of data to obtain and accumulate in the field of autonomous driving, so it is often used in pure vision systems.

3. Development history of end-to-end technology

Between 2016 and 2018, some companies such as NVIDIA and Waymo conducted early explorations, but these results did not achieve mass production. The main reason was that the network architecture at the time was relatively simple, based on CNN, and the learning methods such as imitation learning and reinforcement learning were also relatively basic, lacking advanced network structure support.

The second stage of development is the application of underlying technical architectures such as large language models in this field. The widespread application of network technical architectures such as transformers, as well as the improvement of the complexity and reliability of autonomous driving models, are the core changes in this stage. At the same time, generative AI and large language models have proved that the general API paradigm can lead to physical world AGI, attracting more researchers and practitioners to invest in this direction.

4. Three driving forces behind the attention paid to end-to-end technology in the past year

First, the performance indicators of Tesla's V12 system have increased exponentially. Second, the product value of the end-to-end technology itself, including solving more long-tail scenarios and improving the anthropomorphic aspects of user experience; finally, it is technology-driven, the influence of large language models, and the data-driven characteristics and scaling laws of the end-to-end system.

The impact of end-to-end technology on the organizational structure of car companies and autonomous driving companies is that it will simplify the organizational structure and improve the efficiency of development iteration. At present, many leading OEMs in the industry have officially announced plans to mass-produce end-to-end solutions, although the real end-to-end system may not be mass-produced until next year.

Autonomous driving system and algorithm companies, such as SenseTime, Pony.ai, and Horizon Robotics, have already started end-to-end R&D investment, and expect to have OEM projects in the next six months to a year. At the same time, generative AI and tool chain companies, such as Guanglun Intelligence and Jijia Technology, have also made initial solutions investment and exploration in the end-to-end field.

5. End-to-end technology will lead to further development of upstream technology in the industry chain

End-to-end technology will trigger further development of upstream technologies in the industrial chain and bring about changes in the industry ecology.First of all, as the complexity of AI models increases, the open source ecosystem will play an increasingly important role.The open source community has advantages in talent aggregation, large-scale collaboration, and complex model development, which has been proven in the development of BEV over the past three years. By comparing the performance of open source projects and edge projects from 2021 to 2024, it shows that the open source ecosystem and edge projects are moving forward hand in hand in promoting the development of the field.

Secondly, end-to-end technology poses challenges to traditional simulation and test verification methodologies.Since the vehicle needs to be controlled end-to-end, the open-loop test of the traditional simulation method is no longer applicable, and the fidelity of the existing simulators is problematic. Therefore, a new closed-loop simulation tool chain needs to be developed, which may require the combination of capabilities in the field of generative AI and synthetic data, which will provide opportunities for new companies.

Third, end-to-end technology will drive innovation in chip architecture.The end-to-end AI models with larger parameters and new neural network operators pose new challenges to chip companies, requiring more flexible chip architectures to adapt to the rapid evolution of autonomous driving AI algorithms.

From the perspective of industry development trends, end-to-end technology will accelerate the overall penetration rate of autonomous driving. It is expected that the advent of end-to-end technology will make functions such as highway NOA and urban NOA more popular in the next two to three years and provide services to ordinary consumers.

Due to the strong generalization ability of end-to-end technology, more applications of autonomous driving across geographical regions or countries may appear in the future. At present, autonomous driving companies need to go through a lot of adaptive debugging when entering new countries, but end-to-end technology is expected to reduce this process, improve the ability of cross-scenario applications, and may even bring about the evolution of different scenarios for commercial vehicles and passenger vehicles.

6. The relationship between end-to-end autonomous driving and general-purpose robots

The two fields have been learning from each other and growing together throughout history. Many technologies in the field of autonomous driving originated from the robotics industry, including sensors, perception and positioning algorithms, SLAM (simultaneous localization and mapping), planning algorithms, as well as operating systems and middleware.

In the past five years, although the development of the robotics industry has been slower than that of the autonomous driving industry, the autonomous driving industry has experienced an accelerated industrialization, and the reduction in the cost of hardware components and sensors has actually fed back to the robotics industry. In the past year, the data-driven AI approach represented by end-to-end autonomous driving has gradually been put into mass production, and it is expected to feed back to the general robotics field in the future, which also requires general intelligence.

Looking ahead, autonomous driving and robotics are the two most important areas for achieving physical AI. Many researchers divide general intelligence into autonomous driving and robotics. From the perspective of achieving physical AGI, the autonomous driving industry currently has a more obvious advantage because its task complexity is relatively low and a complete link of acquiring data and iterating algorithms has been established, which provides a model for the robotics industry to learn from.

The development of end-to-end autonomous driving has not only promoted the advancement of autonomous driving technology, but also brought new opportunities and technical references to the robotics industry. The two will continue to promote each other under the development of AI technology and jointly promote the realization of AI in the physical world.

Questions and Answers

Q1: What is the future development trend of LiDAR and millimeter-wave radar as input sensors?

A: The costs of both LiDAR and millimeter-wave radar are constantly decreasing, in line with Moore's Law. The costs of currently shipped LiDAR and 4D millimeter-wave radar are 2,000 to 3,000 yuan and 600 yuan respectively, with a significant gap. It is expected that in two years, the cost of LiDAR can be reduced to about 1,000 yuan, and the cost of millimeter-wave radar can be reduced to more than 300 yuan. Millimeter-wave radar is highly complementary to optical systems such as LiDAR and cameras. Optical systems may fail in severe weather such as rain, fog, snow and dust, while millimeter-wave radar is more reliable in these environments. In the long run, taking advantage of the advantages of different sensors is the key, especially in L3 and L4 autonomous driving. For the L2 level, low-cost solutions may prefer a combination of vision plus millimeter-wave radar.

Q2: In different segments of the TOB field, what is the current gross profit margin of autonomous driving companies and what is their potential gross profit margin?

A: According to the framework of the TO B company's business model, it can be divided into two different models: selling products and operating. From the early stage of industry development, operating can better test the company's capabilities and prove its strength, because operations need to be responsible for the entire process. Relatively speaking, there is not much discussion value about the gross profit margin of selling products, because pricing does not reflect the value of the entire industry chain. Companies that operate are currently in the stage of turning from negative gross profit margin to positive, with a target gross profit margin of 30% or even 40%, while some scenarios are still in a negative gross profit operation state.

Q3: From the perspective of investment institutions, many people think that the time for layout of autonomous driving has passed. From your perspective of the industry, are there any sub-sectors worth looking at?

A: Looking back at the development of the Internet industry, new opportunities and more detailed division of labor will emerge in the mature stage of any industry. The current autonomous driving industry is at a turning point similar to the Internet industry in 2010 or 2012, indicating that unexpected new business models may emerge in the future. From the perspective of the investment stage, now is a good time for some growth-stage investment institutions to invest in the commercialization inflection point, especially in the mining and urban distribution fields, where the fleet size has achieved a significant increase from hundreds to thousands of units within a year, showing a clear inflection point for commercialization. Therefore, for mid- and late-stage investments, now is a good time to make arrangements.

Early-stage investments should pay more attention to technological changes, especially the development of end-to-end technology, which brings new opportunities at the algorithm and tool chain level, such as closed-loop simulation and end-to-end generative data solutions. In addition to end-to-end technology, the increase in the volume of high-end intelligent driving sensors has also brought opportunities to upstream industries. For example, the significant growth in LiDAR shipments has brought potential markets for components such as chips and lasers with a low degree of localization. With the continuous innovation of technology and the continuous expansion of the market, more early investment opportunities will emerge in the autonomous driving industry, just as Pinduoduo and Douyin appeared in the mature stage of the Internet industry. The future development is full of infinite possibilities.

Q4: What are the requirements for vehicle-side computing power in end-to-end? From the perspective of the computing center, how much investment is needed?

A: End-to-end autonomous driving technology does not directly increase the demand for computing power. However, because it is a data-driven solution that complies with the scaling law, in order to achieve better performance, the industry generally expands the model size, which increases the demand for high-computing power chips. Some companies, such as Ideal Auto, adopt a dual-system strategy, in which a more powerful multimodal large model is used for end-to-end, and the other is a lightweight end-to-end model. Both models are deployed on existing autonomous driving chips. The reasoning speed of the large model is slower, showing that it has a higher demand for computing power. As the scale of the model increases, the training cost and the demand for training computing power also increase. Entry-level end-to-end training may require 1,000 cards equivalent to A100, while the leading companies may have reached a layout of 10,000 cards. Taking Tesla as an example, its large-scale investment in training computing power may be difficult for other companies to catch up. Domestic companies such as Xiaopeng and SenseTime have also made huge investments in training computing power. Overall, the development of end-to-end technology has driven the demand for higher computing power chips and greater training computing power to support the training and deployment of larger-scale models. This not only reflects technological progress, but also reflects the industry's pursuit of high-performance autonomous driving solutions.

Q5: For the companies currently engaged in autonomous driving, is there any chance or opportunity to cross over to the field of humanoid or general-purpose robots?

A: The development of humanoid robots in China is closely related to the launch of the Optimus humanoid robot by Tesla. The Figure team also brings together talents from Apple Car and major North American automakers. This cross-industry talent flow is natural because autonomous driving and robotics technologies are interconnected in many ways.

At present, the problem of humanoid robot itself has not been completely solved, so the large-scale migration of autonomous driving technology to the field of robotics has not yet been realized. According to industry opinions, half of humanoid robot companies in the future may be car companies, and I personally partially agree with this view. This shows that with the development of technology, the integration of autonomous driving and robotics will become more and more in-depth.

Observation from Titanium Capital Research Institute

End-to-end architecture is expected to become the ultimate solution for autonomous driving. Intelligent driving systems are currently generally divided into three modules: perception, prediction, and planning. The end-to-end model integrates the three modules, inputs information from the perception end, and directly outputs results at the execution end. Under the modular technical architecture, the transmission of information will be impaired, the system maintenance is difficult, and it cannot cope with complex road conditions. The end-to-end model does not require programmers to write code to formulate rules, but uses massive data to train the system, giving machines the ability to learn, think and analyze autonomously. Titanium Capital will continue to focus on technology iteration and seize opportunities for technological advancement together with industry and capital partners.