Robotaxi is an important landing scenario for L4 autonomous driving. As the third in a series of reports on intelligent driving, this article focuses on Robotaxi and focuses on the four major issues that the market is concerned about: 1) What special requirements does Robotaxi have for the technology stack, and what is the maturity of the current technology? 2) How is Robotaxi's unit economy (UE) model calculated, and what is the room for future improvement? 3) From a policy and corporate perspective, what step is the implementation of Robotaxi? 4) How do you view Robotaxi's investment logic?
summary
What are Robotaxi's special requirements for the technology stack, and how mature is the current technology? The core of the establishment of the Robotaxi model is that the system automates driving, with almost no manual intervention. The requirements of its technology stack are consistent with the strong generalization capabilities and highly anthropomorphic intelligent driving models that urban NOA pursues, and also need to meet safety and redundancy. Currently, Robotaxi has shown significant technological progress, but there is still room for improvement. CICC Research believes that cutting-edge advances in artificial intelligence such as Transformers, end-to-end, multi-modality, and world models are expected to accelerate the upgrading of intelligent driving capabilities.
How is Robotaxi's unit economy (UE) model calculated, and how is there room for future improvement? As a passenger transportation service, Robotaxi mainly charges based on mileage; on the cost side, Robotaxi companies need to bear the costs of vehicle depreciation, safety personnel costs, electricity, parking fees, and insurance premiums. Among them, revenue, depreciation, and safety officer costs are three important factors affecting the Robotaxi UE model. Robotaxi's business model is diverse, and is divided into different forms such as heavy assets and light assets, but in the early days, its UE model was usually at a loss, and future improvements require continuous attention to changes in technology, policies, and fleet size.
From a policy and corporate perspective, what step is the implementation of Robotaxi? China's Robotaxi is gradually entering a transition period from R&D testing to commercial deployment. A number of national and local policies help the industry develop healthily, and relevant overseas policies and regulations are constantly being promoted. From an enterprise perspective, various leading Robotaxi companies in China and the US are continuing to expand their commercial footprint. Currently, their size is relatively limited, and the stages range from road testing to commercial operation.
What do you think of Robotaxi's investment logic? The long-term development of the Robotaxi industry depends on the three elements of technology, commercialization, and policy to form a positive cycle. In the positive cycle, RoboTaxi's UE level is gradually improving, and companies are willing to launch more vehicles, while the expansion of the fleet size can continue to dilute early R&D costs, improve corporate profits, and push the industry towards healthy and sustainable development. However, before a positive cycle is formed, Robotaxi may bear losses for a period of time. According to CICC Research, financial strength, closed data loop, software and hardware integration and technological innovation capabilities are key competitive factors for Robotaxi companies at this stage.
risks
Technology exploration is slow; commercialization is blocked; industry competition is intensifying; policy progress falls short of expectations.
body
Question 1: What are the special requirements of Robotaxi's technology stack, and what is the maturity of the current technology?
Current: Robotaxi requires generalization, personalization, and redundancy of the system, and the existing technology stack still has room for improvement
Robotaxi is an important commercial implementation scenario for L4 autonomous driving. The core of its model is that the system automates driving, with almost no human intervention required. It is a product of artificial intelligence technology developing to a certain level. From the perspective of an autonomous driving company, to achieve the goal of system automated driving, at least 2 conditions must be met:
1) The system's ability is sufficient to achieve automated driving. From the perspective of C-end passengers, they are willing to use Robotaxi. The basic requirement is that Robotaxi is on the same level as the traditional model in terms of safety, comfort, and efficiency. This involves the ability of the system itself, which falls within the technical category discussed in this chapter; 2) The implementation method is cost-effective and conforms to the business logic of the market. Only in this way can autonomous driving companies develop sustainable development, and the price paid by C-end passengers has room to decline. This involves the issue of the unit economy (UE) model. Detailed discussion in a chapter.
Specifically, they break down the dimensions of safety, comfort, and efficiency. Technically, they mean: 1) Safety: The system can avoid or reduce situations such as drunk driving, inattention, and slow response, but the system has a probability of failure, and the neural network itself is also a probabilistic model. Robotaxi needs to ensure the bottom line of safety even if software and hardware fail;
2) Comfort: The system needs to be highly anthropomorphic, handling various scenarios like a human driver to reduce sudden braking and setbacks; 3) Efficient: Robotaxi is required to achieve a certain driving speed, give passengers the flexibility to freely choose routes and places to get on and off the bus, and be able to properly handle corner cases (long-tail scenarios) and interactive games with other cars, rather than simply parking and waiting.
Figure 1: The upgrading of intelligent driving systems places higher demands on hardware and software capabilities
Source: School of Vehicles and Transportation, Tsinghua University, Milimo Zhi Xing's official account, Jiuzhang Smart Driving, CICC Research Department
It can be seen from this that Robotaxi's requirements for the technology stack are consistent with the strong generalization capabilities and highly anthropomorphic intelligent driving models currently being sought by urban NOA, and since there is basically no human intervention, Robotaxi needs to push these dimensions to the extreme and meet safety and redundancy.
Therefore, in Robotaxi's technology stack, on the one hand, we can see higher neural network penetration, more complex model architectures, more detailed model accumulation, and even larger model scale; on the other hand, it will have more sensors (such as lidar) and high-precision map redundancy in terms of perception, and usually add rule code backbone to make decisions to increase the redundancy of the technology stack. In the first report in the series [1], CICC Research summarized these requirements as “completeness” and “redundancy.”
In recent years, in road tests and demonstration operations of various types of L4 at home and abroad, Robotaxi has shown obvious technological progress. It runs smoothly overall, can perform well in turning, changing lanes, and being courteous to pedestrians. It also performs well in complex scenarios such as extreme weather, morning and evening peaks, and urban villages.
However, at the same time, CICC Research believes that the current technical level also has room for further improvement. For example, the handling of some long-tail scenarios (woven bags on the road) and emergency situations (temporary road diversion) is not perfect, driving strategies are conservative (when other cars are encountered, they may directly stop rather than play games), and flexibility needs to be enhanced (in the online car-hailing scenario, drivers can flexibly choose routes and pick-up locations, but Robotaxi has not yet been implemented). Security (including cybersecurity) needs to be continuously strengthened.
According to CICC Research, the existence of these problems may mean that the entire autonomous driving industry will continue to move in the direction of generalization, personification, and safety of models/algorithms.
Figure 2: What the current Robotaxi technology stack can do and where there is still room for improvement
Future: Advances at the cutting edge of artificial intelligence are expected to help improve the Robotaxi technology stack
Returning to the perspective of the technology stack, after years of development, the intelligent driving technology stack represented by Robotaxi has a long history. The classic technology stack usually has a modular architecture that performs various tasks of perception, prediction, decision making, and control, and the AI model penetrates various modules to varying degrees. Under the classic technology stack, some of the problems of intelligent driving have been solved well. For example, the BEV + Transformer successfully solved the vector space construction problem of the sensing module. The introduction of AI models in regulation and control modules can significantly improve ride comfort and bring the ability to drive intelligently into more complex scenarios.
However, intelligent driving technology still has problems to be overcome:
► Perception: Corner case perception is one of the main difficulties, such as strange long-tail obstacles on the road, particularly extreme weather, and temporary road changes. Perception is the first “level” of the intelligent driving technology stack. If the sensing module mistests, then subsequent modules may be under greater pressure.
► Prediction: First, interaction. The future trajectory of a dynamic obstacle is related not only to itself, but also to its interaction with other dynamic obstacles. Predicting complex scenarios usually involves modeling multi-agent interactions. Next is the Corner case, which predicts the future trajectory of various long-tail obstacles.
► Decision: The results output from the perception and prediction module have probability values. The decision module essentially needs to work in an uncertain environment (of course, using high-precision maps, it is possible to sense the uncertainty of the results to a certain extent). In complex scenarios, how to play and interact with other vehicles like a skilled human driver, how to handle various complex operating conditions and even corner cases (such as in the form of a heavy snowy day), and how to make the ride experience of autonomous driving decisions consistent with the human driving model are the main difficulties faced by the decision module.
► Control: Control is more of an engineering and math problem, but in the Robotaxi scenario, hardware tuning, software and hardware coupling may affect the accuracy of control, and ultimately affect the actual driving effect.
Faced with the above technology stack problems, CICC Research believes that cutting-edge advances in artificial intelligence, such as large models, may be useful in the future, and are expected to help upgrade intelligent driving capabilities. Specifically, Transformers, end-to-end, multi-modal, and world models are a few types of cutting-edge developments worth watching.
The Transformer architecture is the cornerstone of large models. It is good at modeling long-distance relationships. It can accept learning methods in the form of prompts, and can effectively link multiple modes of information and merge it into a unified signal. Performance is usually greatly improved as the number of parameters expands. In addition to the field of perception, it is also gradually being used in the field of prediction and decision-making. End-to-end, independent algorithm modules are incorporated into a completely microscopic unified model framework. The feature space can be completely propagated in all parts of the model, and more thorough data driving is possible. Capabilities such as contextual learning, zero-sample learning, logical reasoning, and common sense judgment that have emerged from multi-modal grafting big language models are potential ways to improve the generalization capabilities of intelligent driving technology stacks (previous methods were usually actual vehicle collection/simulation to construct corner cases, and then use data for further training of AI models). By making the model predict the evolution of future driving environments based on historical information, the world model has constructed a more efficient training method for learning the prior distribution of massive intelligent driving data, and has the potential to become a foundation model in the field of intelligent driving. [2]
In fact, CICC Research has seen that many Robotaxi companies have begun to apply the idea of big models to enhance their technology stack capabilities. On Apollo Day 2024, Baidu released Apollo ADFM, a large model that supports L4 autonomous driving. It reconstructs autonomous driving based on big model technology. Using the end-to-end idea of implicit transmission and joint training, the model attributes and data drivers of perception and decision making have been significantly enhanced. The company publicly stated that this technology can balance safety and generalization, and safety is expected to be more than 10 times higher than that of human drivers [3].
Currently, many L4 RoboTaxi companies have begun to apply the idea of Transformers, end-to-end, and even the language module Prompt in the technology stack, and the technology stack capabilities are empowered by large models. CICC Research believes that this may be expected to accelerate the maturity of L4 Robotaxi technology.
In February 2024, OpenAI released the video generation model Sora, showing many characteristics of the world model. The market is paying more attention to the application prospects of Sora in the field of autonomous driving. Considering that autonomous driving has strict safety requirements, and Sora's grasp of basic physical interactions in some cases is still inaccurate, CICC Research believes that currently Sora is still difficult to directly use in autonomous driving on the vehicle side.
In the short term, Sora is expected to be applied to autonomous driving simulation, generating new verification scenarios based on video fine-tuning collected from actual vehicles, or generating new synthetic data for autonomous driving model training.
In the medium to long term, CICC Research is optimistic about the potential of the world model represented by Sora, and has built a more efficient training method for learning the prior distribution of massive intelligent driving data. It has the potential to become a foundation model in the field of intelligent driving. Overseas intelligent driving leader Tesla also publicly shared its exploration and long-term vision on related routes at CVPR 2023.
Figure 3: Classic intelligent driving software technology stack
Note: This picture is only a classic intelligent driving software technology stack, compiled by the CICC Research Department based on literature and other public data. The technology stack currently used in the industry may be different. Sources: “Principles and Practice of Driverless Driving” (Machinery Industry Press, 2018), “Introduction to Autonomous Driving Technology” (Tsinghua University Press, 2019), “Autonomous Vehicle Environmental Perception” (Tsinghua University Press, 2020), “Autonomous Driving Vehicle Positioning Technology” (Tsinghua University Press, 2019), “Autonomous Driving Decision and Control” ( Tsinghua University Press, 2019), “Fundamentals of Autonomous Vehicle Platform Technology” (Tsinghua University Press, 2019), “Self-driving cars: a survey” by Claudine Badue, Ranik Guidolini, etc., Waymo, Tesla AI Day, Xiaopeng's official website, Qingzhou Intelligent Aviation website, CSDN, Zhizhi Auto, Jiuzhang Smart Driving, Smart Car Robot, CICC Research Division
Figure 4: Overview of the big model of intelligent driving
Source: A. Vaswani, N. Shazeer, et al., “Attention is All You Need”, 2017, Y. Hu, J. Yang, et al., “Planning-oriented Autonomous Driving,” 2023, C. Li, Z. Gan, et al., “Multimodal Foundation Models: From Specialists to General-Purpose Assistants”, 2023, X. Wang, Z. Zhu, et al., “DriveDreamer: Towards Real-World-Driven World Models for Autonomous Driving”, 2023, A. Kirillov, E. Mintun, et al., “Segment Anything”, 2023, CICC Research Division
Figure 5: Application prospects of world models such as Sora in the field of autonomous driving
Question 2: How is Robotaxi's unit economy (UE) model calculated, and what is the room for future improvement?
In this chapter, CICC Research will try to sort out the computational framework of the Robotaxi unit economy (UE) model, analyze the main factors affecting its UE model, and sort out the room for improvement of future UE models, so as to discuss from a commercial perspective what aspects are needed to achieve autonomous driving in a cost-effective manner. (For detailed UE model calculations, please refer to the report)
On the revenue side, as a passenger transportation service, Robotaxi mainly charges taxi fees based on the number of kilometers traveled; on the cost side, Robotaxi companies need to bear vehicle depreciation, electricity, parking fees, and insurance premiums. At the same time, in order to help Robotaxi get out of trouble, they usually also need to deploy a certain number of safety personnel in the car or remotely to supervise, so Robotaxi companies also pay for safety personnel.
Revenue, depreciation, and security personnel costs are three important factors affecting the Robotaxi UE model. The detailed analysis of CICC's research is as follows:
► Revenue: Influencing factors include daily order volume, average mileage per order, unit price per kilometer, etc. These factors are closely related to Robotaxi's own technical level and riding experience, consumer acceptance, the scope of operation permitted by the policy, and the number of Robotaxi fleets. CICC Research believes that future improvement paths include: 1) refining and improving Robotaxi's technology stack capabilities to improve safety, comfort, and efficiency attributes to increase consumer acceptance; 2) policies allow the gradual expansion of the area where Robotaxi operates; 3) Robotaxi companies increase the number of vehicles installed in the region, so that the fleet can reach a certain density, shorten the waiting time required for passengers, and form scale advantages and network effects.
► Depreciation: Unlike ordinary vehicles, Robotaxi has added autonomous driving kits such as lidars, cameras, and in-vehicle inference chips. In addition, due to safety and redundancy considerations, Robotaxi usually uses high-precision maps and sets redundancy at the level of online control, sensors, and domain control. All of these will make Robotaxi's vehicle cost higher than that of ordinary vehicles, leading to higher depreciation costs. In the future, CICC Research believes that improvement paths include: 1) Changing from rear-loading to front-loading. This is related not only to the stability of vehicle operation, but also to the cost.
2) Increase the number of fleets and form a scale effect, so there is more room for bargaining when buying bare cars from OEMs; 3) Improved software capabilities and reduced hardware configuration requirements, such as reducing the number of lidars from 6 to 3, and from requiring the use of mechanical radar to semi-solid state radar; 4) Reducing the cost of lidar and other components themselves. The popularity of NOA intelligent driving drives the increase in the usage of various components and drives down prices. CICC Research believes that L4 Robotaxi will also benefit from this cost reduction trend.
► Safety Officer Costs: According to the “Autonomous Vehicle Transportation Safety Service Guidelines (Trial)” issued by the Ministry of Transport in December 2023, the ratio of remote safety personnel to vehicle must not be less than 1:3 [4] for fully autonomous vehicles engaged in taxi passenger transport. In the future, as algorithm capabilities advance and generality increase, the number of scenarios where Robotaxi requires the help of safety personnel may gradually decrease; and continued progress in technology is also expected to provide more confidence in easing policies for people and vehicles than requirements.
As can be seen from the above, the Robotaxi UE model was generally at a loss in the early stages, and future improvements require continuous attention to changes in technology, policy, and fleet size. At the same time, the above factors will also interact. For example, improving bicycle UE will increase Robotaxi companies' willingness to increase the number of fleets, while increasing the number of fleets itself will improve bike UE. According to ARK Invest's research, the single-mile cost of a passenger car is expected to drop to $0.25 with the large-scale implementation of autonomous driving.
It is worth noting that the only business model discussed in CICC's research above is a business model where Robotaxi companies own their own fleets, which is an asset-heavy model. In fact, intelligent driving companies are also experimenting with more diverse business models, such as the third-party asset-holding model (asset-light), the C-end consumer asset-holding model (asset-light), and the joint venture model. Under these models, RoboTaxi's UE model may change.
However, regardless of the business model adopted, CICC Research believes that the core of Robotaxi UE will still focus on technology, policy, and fleet size to raise revenue levels and reduce existing costs.
Figure 6: Illustration of the typical business model of Robotaxi
Question 3: From a policy and corporate perspective, what step has Robotaxi been implemented?
Policy perspective: Multiple national and local policies help the healthy development of the Robotaxi industry
Domestic situation
Overall, China's Robotaxi has gradually moved from R&D testing to a transition period of commercial deployment. The government has successively introduced a series of supporting and regulatory policies and regulations around the implementation of intelligent connected vehicles and autonomous driving functions. At the support level, encourage demonstration applications in specific scenarios, allow open pilot projects, and actively improve software and hardware support; at the regulatory level, pre-entry, in-event management, and subsequent division of rights and responsibilities for intelligent connected vehicles are restricted. In addition, the current policy is also exploring the division of rights and responsibilities for autonomous driving traffic accidents, and some cities have entered the pilot exploration stage.
In June 2024, the first batch of consortia composed of 9 automobile manufacturers and 9 users in China will launch intelligent connected vehicle access and road access pilot projects in 7 cities including Beijing. CICC Research believes that the pilot will support the revision of relevant laws, regulations and technical standards, improve production access and road traffic safety systems for intelligent connected vehicles, and promote the implementation of autonomous driving.
In various cities, relevant policies have also been introduced one after another, gradually expanding the scope of the pilot to help the development of intelligent connectivity systems. According to data from the Ministry of Industry and Information Technology, by the end of 2023, a total of 17 national test demonstration zones, 7 vehicle networking pilot zones, and 16 pilot cities for the collaborative development of smart cities and intelligent connected vehicles had been built across the country. According to CCTV [5], by the end of April 2024, China had opened more than 29,000 kilometers of intelligent connected vehicle test roads, issued more than 6,800 test demonstration licenses, and the total road test mileage exceeded 88 million kilometers. Below, CICC Research sorts out policies related to autonomous driving according to the city dimension:
► Beijing: At the end of April 2022, the Beijing Intelligent Connected Vehicle Policy Pilot Zone took the lead in issuing the first batch of “Unmanned Demonstration Application Road Test” notices. Fourteen unmanned passenger cars from Baidu and Little Ma Zhixing were approved for road testing [6]. As of March 2024, the Beijing High-Level Autonomous Driving Demonstration Zone has issued road test licenses to 29 test car companies. The autonomous driving test mileage exceeds 25 million kilometers [7], covering Tongzhou, Shunyi, Yizhuang, Daxing and other regions. In July 2024, Beijing issued the “Beijing Autonomous Vehicle Regulations (Draft for Comments)”, which emphasizes the need for safety officers to be equipped in accordance with relevant national regulations, and clarifies the determination of responsibility for traffic accidents (where a vehicle causes personal injury or property damage caused by a road traffic accident while the autonomous driving system function is activated, the autonomous vehicle party is responsible, and the vehicle owner and manager are liable for compensation).
► Shanghai: Since 2018, Shanghai has opened autonomous driving test roads in stages and batches, creating four demonstration zones including Jiading, Lingang, Fengxian, and Jinqiao. As of March 2024, 32 enterprises and 794 vehicles in Shanghai have obtained autonomous driving road tests, demonstration applications, and demonstration operation licenses. The cumulative test mileage is about 22.9 million kilometers, covering administrative regions such as Pudong New Area and Jiading, with a total test time of about 1.22 million hours [8]. In March 2024, Shanghai will open the entire Jinqiao Economic and Technological Development Zone and roads such as Shenjiang Road, Hunan Highway, and Lianggang Highway in Pudong as test roads for autonomous driving [9].
► Guangzhou: In 2018, Guangzhou allowed some autonomous driving companies to “try first”, and Robotaxi was launched in Guangzhou. In February 2024, Guangzhou introduced the “Guangzhou Intelligent Connected Vehicle Innovation and Development Regulations (Draft for Comments)”. By July 2024, Guangzhou has opened a total of 827 autonomous driving test roads, with a unidirectional mileage of about 1,666 kilometers, involving 6 administrative districts including Nansha District and Huangpu District. Fifteen companies have carried out various levels of road tests and L3/L4 high-level testing activities in Guangzhou. Autonomous driving companies such as Wenyuan Zhixing, Xiaoma Zhixing, Baidu Apollo, and Didi have cooperated with operators such as Guangzhou Public Transport Group and Qichen Technology to carry out demonstration operations and commercialization demonstration applications [10].
► Shenzhen: In 2022, Shenzhen took the lead in legislating for intelligent connected vehicles and issued the “Regulations on the Administration of Intelligent Connected Vehicles in the Shenzhen Special Economic Zone”. In July 2023, Shenzhen issued the country's first L3 high-speed highway test license plate (BYD); in August of the same year, Shenzhen issued 10 local standards for intelligent connected vehicles. As of May 2024, a total of 944 kilometers of test demonstration roads have been opened throughout Shenzhen, covering various administrative districts such as Nanshan District and Bao'an District. A total of 1,037 road test and demonstration application notices have been issued to 19 enterprises for 349 intelligent connected vehicles [11].
► Wuhan: By the end of 2023, the total open test road mileage in Wuhan had exceeded 3,378.73 kilometers (one-way mileage), covering 12 administrative districts, with a radiation area of about 3,000 square kilometers, reaching a population of over 7.7 million, leading the country in terms of open mileage and number of open areas [12].
► Other cities: In addition, Changsha, Hangzhou, Nanjing and other places are also actively promoting the implementation of Robotaxi: in April 2020, the Changsha Robotaxi service was fully opened to the public, with an operating range of about 130 square kilometers [13]; in May 2024, the “Hangzhou Intelligent Connected Vehicle Testing and Application Promotion Regulations” were officially implemented [14]; Nanjing is building intelligent connected vehicle application scenarios in Jiangning Development Zone, Qinhuai Baixia High-tech Zone, Jianye District Science and Technology Ecological Island, and Lishui Development Zone [15].
Figure 7: Autonomous driving test and operation area display in various cities (partial)
Note: 1) This map only shows the intelligent driving operation area; 2) The Shanghai Intelligent Driving Operation Scope Map only selects the driverless intelligent connected vehicle demonstration section of the section between Nanhui Xincheng and Pudong Airport in Shanghai, which does not represent the scope of intelligent driving operation in the city; 3) The Shenzhen Intelligent Driving Operation Scope Map shows the first batch of pilot situations in 2018, which does not represent the current situation; 4) The Hangzhou Intelligent Driving Operation Scope Map only lists the urban situation, which does not represent the scope of operation of intelligent driving in the city; 5) The scope of operation of intelligent driving in Changsha and Wuhan only lists the scope of operation of Radish Fast Running in July maps, It does not represent the scope of intelligent driving operation of all brands in the city. Sources: Yizhuang Xincheng Official Account, Beijing Shunyi Official Account, Shanghai Communications Official Account, Shangcheng Public Account, Shenzhen Dream Public Account, Guangzhou Intelligent Connected Vehicle Demonstration Zone Operation Center, Guangdong Intelligent Connected Vehicle Innovation Center, Radish Express App, CICC Research Department
Overseas situation
Overseas markets are also making steady progress in formulating policies around autonomous driving. In 2012, the US state of Nevada issued the world's first test license for Google's autonomous vehicles [16]. In 2017, Germany passed the “Road Traffic Law (Eighth Amendment)” and became a country legislating for autonomous driving [17]. In the same year, Germany also published the world's first ethical guide for autonomous driving, establishing 20 ethical guidelines for autonomous driving [18]. In 2021, Germany passed the Autonomous Driving Law to regulate L3 and above autonomous vehicles for the first time [19]. In 2023, an amendment to Japan's Road Transport Vehicle Law officially came into effect, allowing L4 class autonomous vehicles to drive on Japanese roads [20].
Overall, the US has a relatively long timeline for advancing autonomous driving, and has gone through many rounds of iterations. Since 2016, the US Department of Transportation has continuously released Automated Vehicles (AV) 1.0 to 4.0 and the Autonomous Driving Comprehensive Plan (AVCP), providing top-level guidance for the autonomous driving industry, formulating intelligent transportation development strategies using the five-year plan as a blueprint, and exploring policy packages for commercializing autonomous driving.
The specific implementation process varies from region to region in the US, but the overall Robotaxi operating layout continues to expand. Currently, several cities such as Atlanta, Dallas, Houston, Miami, Seattle, Boston, and Pittsburgh have begun pilot operations of driverless cars to carry passengers. Cities such as Phoenix, San Francisco, and Los Angeles have allowed Robotaxi to operate for a fee. Turning the perspective to the specific case of California, the California Vehicle Administration (DMV) and the Public Affairs Commission (CPUC) are advancing exploration of California's autonomous driving road testing and commercial application policies year by year. Currently, many companies have obtained autonomous driving tests and deployment permits. Starting in 2022, California will use San Francisco as a pilot to test the possibility of implementing Robotaxi, and in August 2023, it will first liberalize the commercial operation of Robotaxi for a full time period, all regions, and unmanned drivers.
Enterprise perspective: sorting out the implementation status of leading Robotaxi companies
In recent years, the size of the Robotaxi fleet has continued to expand, and the number of mileage and travel orders has increased, but compared with the traditional travel industry, the overall volume is still relatively limited. Currently, first-tier cities are still the focus of the layout of leading domestic Robotaxi companies, and considering the strong support for autonomous driving in cities such as Wuhan and Changsha, CICC Research believes that other cities are also expected to become important areas for Robotaxi implementation in the future. At the level of hardware facilities, there are still a few companies that have achieved mass production of front-mounted units, and more front-mounted mass-produced models are in the process of being developed. In terms of business model, most Robotaxi companies choose to set prices according to the travel service method; in terms of operating hours, Radish Express is one of the few companies (in Wuhan) that have tried to operate 24x7 [21].
Below, CICC Research has briefly sorted out the business conditions and latest developments of some leading Robotaxi companies:
Wenyuan Zhixing
Established in 2017, Wenyuan Zhixing is an L4 autonomous driving company. Currently, Wenyuan Zhixing's product matrix is divided into two parts: L4 and L2+/L3. L4 products cover various scenarios such as online car-hailing, buses, freight transportation, sanitation, etc., including autonomous taxis, autonomous minibuses, autonomous freight vehicles, and autonomous sanitation vehicles; L2+/L3 are mainly high-end intelligent driving products. Wenyuan Zhixing is an autonomous driving company that also has four autonomous driving licenses from China, the United States, the United Arab Emirates, and Singapore. [22]
Wenyuan cooperated with OEMs such as Guangzhou Automobile and Nissan on Robotaxi to implement full-stack software and hardware solutions. In 2019, the company opened an autonomous taxi service in Guangzhou and built the first Robotaxi fleet to land in a first-tier city in the country. In 2022, Wenyuan Zhixing broke through 10 million kilometers of autonomous driving mileage on open roads. In June 2023, the company was approved to launch a fully unmanned robotaxi demonstration application in Beijing, and in November 2023, it was approved to launch a pure unmanned commercial toll operation in Beijing. [23]
My Little Pony Tomoyuki
My Little Pony was founded at the end of 2016 and has three business segments: Robotaxi, RoboTruck, and PoV. Its products cover L2+ to L4 intelligent driving. In December 2018, the company launched PonyPilot, the first autonomous driving travel service in China, which aims to provide public travel services through self-operated or cooperative L4 fleet operations. [24]
Currently, Little Pony Smart's autonomous travel services cover Beijing (Yizhuang, Daxing Airport), Shanghai (Jiading), Guangzhou (Nansha), and Shenzhen (Qianhai, Bao'an, Nanshan). Currently, the company provides fully driverless vehicle services for passengers in Beijing and Guangzhou, and “unmanned driving seats” in Shanghai and Shenzhen. As of May 2024, My Little Pony has accumulated more than 33 million kilometers of autonomous driving road test mileage, of which unmanned autonomous driving test mileage exceeds 3 million kilometers. [25]
In close cooperation with Toyota, mass production of the Robotaxi front assembly was arranged. On April 25, 2024, Little Pony Zhixing, Toyota China, and Guangzhou Automobile Toyota announced at the Beijing Auto Show that the tripartite joint venture will complete registration. The company plans to launch a thousand-scale Platinum 4X autonomous vehicle in the first phase. After the production line is offline, it will be seamlessly connected to the Little Pony Zhixing Robotaxi operation platform to develop large-scale fully driverless travel services in first-tier cities in China. [26]
Baidu
Baidu's layout for autonomous driving can be traced back to the Apollo Program released in 2017, and its autonomous driving product Radish Run was first launched in 2021 [27]. Since its development, Radish Express has achieved a wide scale of implementation and commercial operation. It has obtained driverless test licenses in 11 cities including Wuhan and Beijing. Among them, Wuhan, Beijing and other places can operate for a public fee. According to Create 2024, Radish Express's fleet size in Wuhan is expected to exceed 1,000 units in 2024 [28].
According to Leifeng Network [29], take the 5th generation Radish Express driverless car as an example. The cost of this model is 0.48 million yuan and was customized based on the BAIC Jihu Alpha T pure electric model. At the same time, Baidu is also working with Jiangling to build the next generation lower-cost front-mounted mass-produced model, the Yichi 06 [30], equipped with more than 40 different types of sensors (according to Hesai Technology's announcement [31], its main lidar is exclusively supplied by Hesai. The bike is equipped with 4 ultra-high-definition long-range lidars AT128), a dual-chip configuration, and the vehicle price is about 0.2 million yuan.
autoX
AutoX was founded in 2016. The company's Robotaxi products have now obtained driverless test licenses in Beijing, Shanghai, Guangzhou, Shenzhen, etc., and are operated publicly for a fee in cities such as Shanghai and Shenzhen. The company's fifth-generation Robotaxi model cooperated with Chrysler. The original factory supports car-level redundant wire control, is equipped with 50 sensors, and the vehicle control unit has 2,200 TOPS computing power. [32]
Waymo
Waymo was founded in 2016, and its predecessor was Google's autonomous driving project that began in 2009. Currently, it mainly provides L4 level autonomous driving platforms (Waymo Driver). The business model includes providing Robotaxi services (Waymo One) and driverless cargo transportation services (Waymo Via). Currently, Waymo One has been approved to operate for a fee in Phoenix, San Francisco, etc., with a maximum driving speed of 65 mph. The California Department of Transportation does not set an upper limit on the size of the vehicle fleet [33], and the company operates 24x7 in San Francisco and other places. The Waymo One includes models such as the Jaguar i-Pace [34], as well as the Robotaxi model, which is being mass-produced in collaboration with Kyokrypton. As of December 2023, Waymo has traveled a total of about 20 billion miles in real and simulated environments (of which it has driven about 20 million km on public roads) [35].
Chart 8: Summary of supporting hardware and suppliers for autonomous driving companies
Chart 9: Summary of highlights of the commercialization of autonomous driving companies in 2024
Q4: How do you view Robotaxi's investment logic?
In the first three chapters of this article, CICC Research conducted a detailed analysis of the three key factors affecting the Robotaxi industry: technology, commercialization (UE), and policy. In fact, CICC Research believes that the long-term development of the Robotaxi industry depends on a positive cycle formed by these three elements:
1) After Robotaxi's technology reaches a certain level, the policy opens up some areas for testing and demonstration applications; 2) In the process of field testing and application, Robotaxi companies accumulate more data to train and improve algorithms; 3) As technical capabilities improve, safety, comfort, and efficiency properties are verified, providing more confidence for further policy support.
4) The policy allows Robotaxi to operate on a wider scale, and the fleet size of Robotaxi companies is further expanded. In this positive cycle, RoboTaxi's UE level is gradually improving, and companies are willing to launch more vehicles, while the expansion of the fleet size can continue to dilute the high R&D costs in the early stages, improve corporate profits, and ultimately drive the entire industry to achieve healthy and sustainable development.
Figure 10: The three elements of policy, commercialization and technology form a positive cycle, which together influence the long-term development of the Robotaxi industry
Currently, CICC Research has seen that Robotaxi companies have made some progress in technology and commercialization, and policies also support the healthy development of the industry; however, the three major factors themselves are a cyclical and gradual process, and the development of the industry takes more time and requires a long-term approach.
However, if you want to speed up the operation of the positive cycle, usually one of these factors first obtains breakthrough progress (CICC research suggests that this factor is more likely to be technology), and then pushes the other two factors forward, thus forming a forward flywheel.
CICC Research believes that to achieve this goal, the current industry still has some room for improvement; and before the flywheel is fully operational, Robotaxi companies may bear losses for a period of time.
Based on the above judgment, CICC Research attempted to summarize the key competitive elements of Robotaxi companies at this stage [36]:
► Financial strength: Advances in Robotaxi technology are inseparable from talent and computing power. The support behind it is capital; before the number of fleets reaches a certain size, Robotaxi is usually owned by the enterprise, forming capital; before the industry's flywheel is formed, Robotaxi companies may bear losses for a period of time; the responsibility for traffic accidents is borne by Robotaxi companies. All of the above factors make Robotaxi companies need strong financial support, and what is behind this depends on whether the company has other profitable business hematopoiesis, strong financing capacity, or large cash reserves. At the same time, CICC Research believes that it is also a viable path for autonomous driving companies to take the lead in commercialization in scenarios such as closure and low speed, and obtain certain profit support. For example, in recent Guangzhou road surveys, CICC research can see an increase in the proportion of autonomous driving applications in low-speed scenarios such as microcirculation traffic; in scenarios such as ports and mines, autonomous driving continues to be implemented.
► Closed data loop: As the penetration rate of neural networks in the autonomous driving technology stack increases, a large amount of training data needs to be fed; as AI big model exploration deepens, the number of model parameters may increase further, and Robotaxi companies are required to build mature data engines and tool chains. According to CICC Research, the strength and weakness of autonomous driving companies' data closed-loop capabilities will be directly linked to their fleet size, product richness, model architecture, and engineering capabilities.
► Integration of software and hardware: The tuning and adaptation of models, the stability of vehicle hardware, and electronic and electrical architecture are all closely related to the ability of autonomous driving. For smart driving companies that focus on algorithms, deep cooperation or even binding with car companies is critical, because pre-loading involves the car companies' willingness to cooperate, including whether car companies are willing to redesign the entire underlying architecture, whether they are willing to deeply open up various interfaces, etc. In addition, CICC Research also suggests focusing on the trend of deep chip and software coupling.
► Technological innovation ability: Robotaxi technology still has room for improvement. The current technology stack has not yet converged. Adhering to long-term principles and being able to innovate rapidly is an important soft power. According to CICC Research, on the one hand, technological innovation capabilities require highly qualified software and hardware engineers, and on the other hand, they are also inseparable from the soul figures of the R&D team and the overall corporate culture and management incentive mechanism.
Chart 11: In Guangzhou road tests, the proportion of autonomous minibuses, etc. increased
Chart 12: Autonomous driving solution providers are expected to gain higher industrial value
Finally, given the positive cycle of the industry, CICC Research suggests focusing on investment opportunities in the following segments:
► Autonomous driving solution providers: According to ARK Invest's estimates, autonomous driving platform solution providers are expected to gain higher industrial value by 2030. CICC Research suggests evaluating various autonomous driving solution providers from the dimensions of financial strength, closed loop of data, integration of software and hardware, and technological innovation, and focusing on investment opportunities for leading companies.
► OEMs: It is recommended to focus on leading OEMs with leading positions in the field of autonomous driving. Based on full-stack self-development and fleet size, leading OEMs can often build real-world AI through huge real-world data and evolve into software-driven AI technology companies. CICC Research suggests focusing on the layout and progress of these OEMs in the Robotaxi field.
► Hardware industry chain: At the hardware level, autonomous driving requires hardware support such as domain controllers, redundant wire control, lidar, and inference chips. It is recommended to focus on driving the performance of relevant racetrack hardware manufacturers after the launch of the Robotaxi industry.
[1] For details, please refer to our report “Exploring Smart Driving (1): Disassembling the Intelligent Driving Technology Stack” released on January 29, 2024
[2] For more analysis of the big model of intelligent driving, please see our report “Exploring Smart Driving (2): Empowering Intelligent Driving” on January 31, 2024
[3] Source: https://mp.weixin.qq.com/s/dJDCFRWZLq0aAhhf8xWM_w
[4] Source: https://xxgk.mot.gov.cn/2020/jigou/ysfws/202312/t20231205_3962490.html
[5] Source: https://mp.weixin.qq.com/s/ySyyxz9cKeD4uveyyz_ijw
[6] Source: https://kfqgw.beijing.gov.cn/zwgkkfq/ztzl/cxfbqdkfq/wqhgkfq/wqhg2022/wqhg2022n4y/202204/t20220429_2696369.html
[7] Source: https://new.qq.com/rain/a/20240315A041X900/
[8] Source: https://mp.weixin.qq.com/s/yJR21gxN-dYvO3akIZtD1w
[9] Source: https://www.shio.gov.cn/TrueCMS/shxwbgs/ywts/content/0afc17f1-69ba-4f5e-b933-def729d973f1.htm
[10] Source: https://jtj.gz.gov.cn/gkmlpt/content/9/9744/post_9744164.html#14313
[11] Source: https://mp.weixin.qq.com/s/Xxv5c1glkZG9-IiKIb7ldw
[12] Source: https://3g.wuhan.gov.cn/sy/whyw/202405/t20240531_2410344.shtml
[13] Source: https://mp.weixin.qq.com/s/7VTbnr7Gv4NVieL399t01w
[14] Source: https://www.hzrd.gov.cn/art/2024/4/25/art_1229690462_19702.html
[15] Source: https://mp.weixin.qq.com/s/l_Vx1uPzbxdBe1RKySOLZw
[16] Source: https://edition.cnn.com/2012/05/07/tech/nevada-driveless-car/index.html
[17] Source: https://www.springerprofessional.de/automatisiertes-fahren/unternehmen---institutionen/bundestag-beschliesst-gesetz-zum-autonomen-fahren/12193302
[18] Source: https://mp.weixin.qq.com/s/_qgrigNeXypW-EdzAuLrMQ
[19] Source: https://mp.weixin.qq.com/s/_qgrigNeXypW-EdzAuLrMQ
[20] Source: http://www.news.cn/asia/2023-05/22/c_1129637695.htm
[21] Source: https://mp.weixin.qq.com/s/FLmlojPSQycOXvION8_YYw
[22] Source: https://mp.weixin.qq.com/s/GUnVNM4SV2C6H3vZ7A-EBw
[23] Source: https://mp.weixin.qq.com/s/DW8Ta7kRYxvl0TkWBT9X5g、https://mp.weixin.qq.com/s/sodSZFfT5kazAn-pAB2oVA、https://mp.weixin.qq.com/s/hd0T47umF_7UjrywnRtaeg, https://mp.weixin.qq.com/s/KILpjU6TjFhpwN-CgelM_g
[24] Source: https://mp.weixin.qq.com/s/amHdZZrBtXJQ13eHySD1NQ
[25] Source: https://mp.weixin.qq.com/s/5k7mODy94vZCoOgtCLpYdg、https://mp.weixin.qq.com/s/HZSn9YNV-oq8ZWEWZX1fnA
[26] Source: https://mp.weixin.qq.com/s/j5b0nzCLDd-rhaiu_nPJpQ
[27] Source: https://mp.weixin.qq.com/s/dgVAWFzOgTm5Dd1XIwcMgQ、https://mp.weixin.qq.com/s/dJDCFRWZLq0aAhhf8xWM_w
[28] Source: https://mp.weixin.qq.com/s/QyFKud4d4L69lDiQS5u-vQ
[29] Source: https://mp.weixin.qq.com/s/MbCpSUO5L43lUYSF2PXkoA
[30] Source: https://mp.weixin.qq.com/s/WDbEUPr9EofqyAMRj_ruDQ
[31] Source: https://mp.weixin.qq.com/s/WDbEUPr9EofqyAMRj_ruDQ
[32] Source: https://mp.weixin.qq.com/s/GIpqL3RcNlRggTBIfhZu2g、https://mp.weixin.qq.com/s/IwVI2v8EstPJJHISCvxAHg、https://mp.weixin.qq.com/s/o47TOORP0mWirAvGov-Xqg, https://mp.weixin.qq.com/s/ycwlPptEpSyM9VU68zxgVQ、https://mp.weixin.qq.com/s/lb09tiX4PBeGh_OAVFGXiA
[33] Source: https://mp.weixin.qq.com/s/K2dlHzsCBODkqAEjoc_Bow
[34] Source: https://mp.weixin.qq.com/s/QT0NmtQ_FWJT8VnEd_ktmQ
[35] Source: https://mp.weixin.qq.com/s/r7Asbu83-sFo99ummAOUSw
[36] For more analysis of competitive factors in the entire intelligent driving industry, please refer to our report “Exploring Smart Driving (2): Empowering Intelligent Driving” on January 31, 2024: AI Big Model Wave Empowers Intelligent Driving”
Editor/Jeffrey