Report l Special Research on Autonomous Driving: Facing the Timetable for the Implementation of L3 and L4 Levels

◼ Technical difficulty: The large-scale application of L2 ADAS is based on Tier1

Based on powerful systems, modular capabilities and automotive-grade products. L3 level is a real watershed in terms of driving safety risks and technical difficulties, and is the focus of recent mass production cooperation for passenger car OEMs . Some first-tier market companies focusing on passenger car scenarios already have L3 level mass production capabilities. L4 level is of great significance to cost reduction and efficiency improvement in commercial scenarios , which can greatly reduce labor costs. In the later stage, there is no need to configure safety officers in the car , but the implementation time in China is affected by uncertain factors.

◼ Scenario implementation and technical difficulty: The passenger car scenario is currently undergoing a transition from L2 to L3, and mostly adopts a perception solution of lidar + millimeter-wave radar + camera, using FPGA or partially ASIC chips , combined with high-precision maps in scattered areas for decision-making and control; commercial vehicles are mainly used for high-speed freight, and implementation can quickly reduce costs and improve efficiency. The L4 level is the main breakthrough direction, and mostly adopts camera vision + lidar perception solutions, and GPU chips as the computing center. There are many participants but few companies with mass production cooperation.

◼ Industry KSF (Key Success Factors):

✓ Algorithm companies (providing solutions and systems): order delivery capabilities (whether there is real order cooperation with OEMs and Tier 1 for products), personnel background of the R&D team (time invested in researching relevant models and algorithms), whether there are mature products for testing and operation, whether the sensor combination cost of relevant scenarios is suitable for mass production, and whether the strategic planning is stable (focusing on the time for the implementation of target scenarios)

✓ Hardware supply (sensors and chips): Millimeter-wave radar is mature and is the solution for most algorithm-based companies.

The main equipment of the perception layer; the automotive-grade laser radar has a long development cycle, and the purchase cost is high for testing, and it must be exposed outside the car for perception, so it is relatively rarely involved in passenger car solutions; camera vision solutions are mostly used in L2 level, and the application scale of L3 level and above is large; chip development in the future tends to be ASIC dedicated chips, and currently GPU general- purpose chips are more used in commercial scenarios suitable for L4 level and above.

1. Autonomous driving: timetable postponed, "geo-fencing" restricted scenarios implemented

➢As the traffic ceiling of mobile Internet gradually reaches its peak, the digital integration of the Internet with traditional industries such as agriculture, industry, construction and services will become a new trend. The combination of industrial Internet with technologies such as 5G and cloud computing will accelerate the transformation of the real economy.

➢As an indispensable smart mobile device in the industrial Internet scenario, cars will create a replicable and cyclical business model closed loop in combination with different landing scenarios as the new generation of automobile technology revolution such as new energy, intelligent networking, and autonomous driving innovations.

➢The draft of the "Intelligent Connected Vehicle Innovation and Development Strategy" released by the National Development and Reform Commission plans to achieve a 50% proportion of smart new cars in 2020. As one of the most important technical links of smart new cars, autonomous driving is also constantly undergoing market applications and trial operations at different levels.

➢At the beginning of the surge in popularity of the autonomous driving industry, algorithm companies and OEMs planned the implementation of L4~L5 levels of autonomous driving between 2018 and 2022. However, considering the government's open attitude towards autonomous driving, the occurrence of complex road emergencies and the adaptability of the "geographical fence" effect to some scenarios, the implementation time of different scenarios varies significantly.

➢Currently, all OEMs and Tier 1s have L2-level mass production capabilities. Algorithm companies are targeting the L3-level passenger car and L4-level multi-commercial scenario markets. Leading Internet companies and large OEMs have released their own autonomous driving-related platforms to further concentrate resources.

1.1 Concept of autonomous driving: perception-decision-control, algorithm is the core of the solution

➢Autonomous driving means that smart cars use sensor equipment installed on the car (including 2D photography visual perception, lidar, millimeter wave radar, etc.) to perceive the driving environment around the car , combine it with high-precision navigation maps and other map data, perform rapid calculations and analysis, continuously simulate and deeply learn potential road conditions and make judgments, and further use algorithms to plan the most ideal or most appropriate driving route and method for the car, and then feedback to the control system through the chip to perform actual operations such as braking and steering wheel control.

➢In summary, autonomous driving can be interpreted in steps, including the perception layer, decision layer, and execution layer. The perception layer uses various visual devices and radars to perceive the surrounding environment, combines chip algorithms and V2X (Vehicle to X) to obtain environmental information, and uses the deep algorithms and rule algorithms contained in the decision layer to continuously simulate road conditions, plan the best route, and feed back to the control layer to implement driving operations.

➢The scope and scenarios of autonomous driving are not limited to passenger cars on urban roads. Many commercial companies have started different paths of technology research and development based on different applicable scenarios and solution directions since their establishment. From 2D photography visual data acquisition to 3D lidar modeling, from passenger cars to commercial vehicles, from complex urban roads to regular and limited scenarios...commercial companies involved in the field of autonomous driving have cooperated with upstream vehicle manufacturers and OEM companies as Tier1/Tier2 to start their own commercialization paths.

1.2ADAS: Assisted unmanned driving, the driver makes decisions with the help of the system

➢ADAS (Advanced Driver Assistance System) is an integrated control system for active safety functions. It uses sensors such as radar and cameras to collect data about the car's surrounding environment, identify and track static and dynamic objects, and make behavioral decisions based on map data, so that the driver can be aware of possible dangers and directly control the vehicle to avoid collisions when necessary, which can effectively improve driving safety and comfort.

➢ADAS is the prerequisite for realizing autonomous driving. Both autonomous driving and ADAS (Advanced Driver Assistance) use sensors to collect data inside and outside the car to feedback abnormal information around the car. The difference is that ADAS feedbacks abnormal information about the surrounding area to the driver, and the driver performs driving operations based on the feedback of road information and sensor data. The highest level of autonomous driving is to transmit the data fed back by the sensors to the decision-making layer for decision-making, and finally the control layer guides the behavior to the system, and the system completes the final operation.

➢ADAS does not have a specific coverage range in its definition, ranging from non-automated to unmanned

Driving technology innovation can be seen as part of ADAS. The implementation process of ADAS starts from hardware equipment operating the perception system, databases, chip algorithms and other planning specific decisions, and motors and other control units operating the control system. The overall process is inseparable from the perception-decision-control operation line . At present, ADAS includes but is not limited to adaptive cruise control, blind spot detection, forward collision warning system, night vision system, etc.

1.3 Autonomous driving classification: L1-L4 has limited applicable scenarios and large differences in implementation time

➢Since there will be different degrees of system intervention in the process from no automation interference to the final unmanned driving, national associations have divided autonomous driving into different levels and standards, and commercial companies in various countries announce the R&D stage and implementation results according to the divided levels. Currently known standards include those formulated by CAAM (China Association of Automobile Manufacturers), NHTSA (National Highway Traffic Safety Administration), and SAE (American Society of Automotive Engineers). The internationally accepted standards are mainly based on the six stages of L0-L5 formulated by SAE.

➢L0: There is no automation equipment involved in this stage. The driver controls the car throughout the whole process.

➢L1: Single function automation. In a specific driving environment, a single assisted driving system can provide feedback to the driver by obtaining information about the vehicle's surrounding environment, but the dynamic operation is completed by the driver.

➢L2: Partial system automation. Multiple auxiliary driving systems operate the car's lateral and longitudinal driving actions simultaneously based on environmental information, and dynamic operations are still completed by the driver.

➢L3: Under certain circumstances, the system completes all dynamic operations, but the driver needs to respond to the system when special circumstances occur. Currently, most commercial companies focus on the implementation of this stage.

➢L4: Under certain circumstances, the system is responsible for performing all dynamic driving actions even if the driver does not respond to special situations.

➢L5: The system performs dynamic driving actions in all road conditions, and the driver can manage the system.

1.4 Autonomous driving scenarios: logistics and transportation are highly commercialized, and urban road conditions are complex

➢ From the concept to the development to the current stage, the two most important goals of autonomous driving are to reduce driving risks and improve safety, and then reduce costs and achieve mass production. Not only are there differences in the models of passenger cars and commercial vehicles, but the scenarios they are applicable to are also quite different, and the business paths are different.

➢ The representative companies here are only those that provide scenario solutions, excluding hardware manufacturers, map providers, vehicle manufacturers and Tier 1. Please refer to the subsequent chapters for detailed analysis.

2. Full “Scene” Tracking: Capital Layout is Concentrated, and Startups’ Commercial Scenarios Are Focused on Logistics

➢This chapter will study and analyze autonomous driving from a macro perspective, exploring the development of the industry under policy influence, the investment layout and cycle changes of investment institutions under capital background , and the impact of vehicle manufacturers, OEMs and Tier 1 on domestic and foreign primary market companies.

➢Among them, in the fourth chapter of this chapter, we will focus on analyzing the current situation in each segmented scenario and the factors affecting the hematopoietic capacity of domestic and foreign companies. Starting from the markets where their respective markets are located, we will analyze the feasibility of each scenario based on market demand, existing risks, and cost reduction and efficiency improvement brought about by mass production of autonomous driving.

2.1 Policy regulations: Road test standards and development strategies have been implemented

➢In January 2018, the National Development and Reform Commission issued the "Intelligent Vehicle Innovation and Development Strategy" (Draft for Comments), which set out a three-stage vision timetable for the innovative development of intelligent vehicles by 2035.

2.2 Capital Feast: Institutions select one company in the commercial, passenger and hardware sectors to follow up on the investment

➢From the perspective of capital participation, autonomous driving, as an industry that has been heavily invested by venture capital, has always attracted attention and controversy due to the huge amount of investment. For example, in February this year, SoftBank Vision Fund invested $940 million in autonomous driving company Nuro.ai, which nearly doubled the amount of financing of Aurora, which raised $530 million not long ago.

➢On the one hand, it is difficult and time-consuming to achieve autonomous driving above L3 level, and the equipment and chips invested in R&D are used for testing and are not mass-produced, so the single investment cost is high; on the other hand, system driving above L4 level will have a great impact on the cost reduction of commercial scenarios and the safety of passenger cars. The valuation of algorithm-based companies will increase rapidly after cooperating with OEMs, and the cooperation of orders will double the revenue. Domestic venture capital institutions are relatively more rational in participating in the investment in the field of autonomous driving, and select one company for tracking in different scenarios and hardware.

2.3 Analysis of landing scenarios: Commercial logistics is progressing rapidly, but the prospects of passenger cars are unpredictable

➢The difficulty of landing commercial vehicles and passenger vehicles is affected by different factors.

➢ Passenger cars: At present, there are no complete legal provisions or policy documents related to the road or mass production of passenger cars. Most solutions cannot meet the technical requirements of L3-L5 levels, plus safety issues, legal liability issues, consumer awareness and other issues. At present, it is difficult to implement, but with the synchronization with the mass production of vehicle manufacturers, there has been a turnaround recently.

➢ Commercial vehicles: Commercial vehicles have different application scenarios, so their implementation varies. Currently, the mainstream scenarios with strong implementation include autonomous driving ride-hailing services, high-speed transportation, port freight, mining areas, municipal sanitation, and last-mile logistics. Subsequent content will also analyze commercial scenarios from the perspective of cost, efficiency, optimization and other data levels.

2.3.1 High-speed transportation scenarios for trunk logistics

➢ Trunk logistics scenarios are mainly based on L4 system control. Shuttle between cities with a large demand for road freight. The pricing method is closely related to the volume, weight, category, and other factors of the goods being transported. The main reason for the implementation of autonomous driving in trunk logistics is nothing more than cost reduction and efficiency improvement.

➢From the perspective of cost reduction, we first consider the components of cost. We analyze the distribution process of cargo transportation, personnel management costs, driver wages, high-speed transportation tolls , excess fuel consumption caused by driving, and potential violation costs under the "fine economy".

➢So how does autonomous driving reduce costs? We start with the business model of commercial companies . The technical route that companies like TuSimple and Embark directly cut into is to skip L3 and go directly to L4, because it is directly linked to costs.

➢ Due to the complexity of V2X in driving scenarios and mass production in cooperation with vehicle manufacturers, the development of L3~L5 based on standards step by step during the research and development process is a safe and stable path.

➢Relying on the commercial vehicle market in different scenarios, the research and development of L3 cannot solve the high driver labor costs and fuel consumption problems. Therefore, going straight to the L4 level of research and development is the fastest way to pursue commercialization.

➢ Costs are quantified by different variables, and we mainly analyze driving costs and management costs.

Driving costs include driver salary + fuel costs and fuel consumption + road and bridge tolls + insurance + fines + other fees , while management costs include vehicle maintenance + training + operator salaries and other expenses.

➢At the same time, considering the logistics costs of China, the United States and Japan, transportation costs are the core costs with the highest proportion including warehousing and management costs. Therefore, how to solve the high transportation costs through autonomous driving is the focus of this section.

➢ Cost: From the perspective of driving costs, there is limited room for reduction in hard costs such as road and bridge fees, which will not be discussed here. Instead, key cost components such as driver salaries, accidents, fuel consumption, and fines are the main points of discussion.

➢ Driver salary: According to CNBC’s report on truck autonomous driving, based on the statistics of the U.S. Bureau of Labor Statistics, the median annual income of truck drivers is about $44,500.

➢The annual salary of a single driver in China is at least more than 200,000 yuan. Combined with the situation where two drivers take turns on duty in a single vehicle, the rising demand for transportation will face a shortage of supply in the context of a small number of trunk logistics drivers and low growth.

➢ Level 4: With Level 4 autonomous driving technology, the driver can sense the surrounding vehicle environment and dynamics through sensors without the need for a driver, and make decisions for the next step of driving based on the location and signs of the structured road. The system will make the final decision. During this process, the driver's workload is greatly reduced. In the future, as policies are relaxed and technology continues to develop, it will be possible to achieve complete unmanned operation without the need for a safety officer in the car.

➢ In terms of costs, taking TuSimple's business model in Arizona and Texas as an example, although the cost of each vehicle in the early trial operation was high (independent purchase of laser radar and other equipment), the subsequent mass production will customize the cost standards according to the needs of different customers. In addition, the mass production of the front-mounted system can reduce the accident rate, improve driving safety, and reduce the investment in driver costs.

➢ Fuel consumption and fines: On domestic highways, commercial transport vehicles of different ages will reach their service life within a certain period of time, but the high cost of purchasing vehicles will still make companies choose to continue using the vehicles. Therefore, during inspections on highways, fines due to non-compliance with emission standards are often encountered.

➢ Emission standards According to statistics from the United States Department of Transportation, truck emissions account for 16% of global greenhouse gases, and 67% of trucks do not use clean fuel. At the same time, because drivers will react to road conditions at any time while driving, driving in unfamiliar and complex road conditions will lead to increased fuel consumption.

➢ Trucks equipped with L4 can perceive structured road information in advance, and the system can operate to minimize fuel consumption, avoiding unnecessary time consumption caused by drivers' repeated driving due to unfamiliar road conditions.

➢From the perspective of management costs, the main thing is to improve the dispatching costs required when assigning drivers to perform operations, including the waiting time for vehicles to travel back and forth and the idle time costs caused by staff shortages.

2.3.2 Semi-enclosed hub scenarios (parks, mining areas, road cleaning)

➢ There are many types of semi-enclosed hub scenarios, which we will focus on here. Municipal sanitation, to be precise, belongs to the commercial scenario under open roads, but because the market space is relatively small, there are fewer participants, and the single coverage of municipal sanitation is small, so it is discussed here together.

➢The scenarios discussed here include mining areas, ports, municipal sanitation within a certain range, logistics parks (including last-mile logistics), etc.

——Municipal sanitation

➢Artificial intelligence companies including Zhixingzhe Technology and Xiantu Intelligence are the main participants in the sanitation scene. Since municipal sanitation scenes are mostly fixed semi-closed areas, the road structure is more complex than that of closed areas. However, in similar areas, the municipal environmental protection vehicles have low requirements for speed and timeliness. Therefore, under the condition of full system control, it can effectively reduce labor costs and improve cleaning efficiency.

➢From the perspective of landing difficulty, most cleaning vehicles are manual transmission and do not have more electronic control equipment, so it is difficult to install them later. Solutions such as Xiantu Intelligent have selected new energy vehicles of 3 tons to 18 tons to install wire control for control. Zhixingzhe Technology's unmanned cleaning vehicle "Wo Xiaobai" has also cooperated with Shougang and put into mass production.

——Mining areas and ports

➢ Mining areas, logistics parks and ports are closed areas with large driving areas, high controllability and stable driving at low speeds. However, they may also be affected by special climate environments such as sea breezes and high temperatures, and the risk factor is slightly higher.

➢ In this environment, despite the harsh environment, the operator has high requirements for the driver's familiarity with the driving route and the cargo in the corresponding scenario, resulting in a greater shortage of drivers whose positions are not very attractive. Autonomous driving has greatly improved the cost reduction and efficiency improvement functions in this scenario. It solves the problem of reduced operational efficiency due to staff shortages and reduces the risk of dangerous accidents.

➢Truck transformation in this scenario is also a problem that algorithmic companies need to overcome. Heavy trucks in mining areas and ports have large load units and a long service life for a single vehicle, so the difficulty of transformation and handling of wire-controlled equipment is huge.

➢ The existing solutions are mainly post-installation, and most algorithm companies are also negotiating pre-installation cooperation with heavy truck OEMs to further optimize costs. However, for existing mining areas and ports, the overall environmental transformation and vehicle pre-installation transformation are still very difficult.

——Logistics Park and Last-mile Delivery Services:

➢The problem that the vehicle transformation of the logistics park solves is that with the annual increase in freight volume, the operation time of transport vehicles is longer, the movement and unit turnover of large goods such as containers in the park are also larger, and parking needs to achieve centimeter-level accuracy and interact with cranes. According to the data of the National Bureau of Statistics, the national freight volume and road freight volume are increasing year by year, and the freight transportation volume has increased from 41.76 billion tons in 2015 to 51.46 billion tons in 2018.

➢ The last mile delivery service is mainly for express delivery and takeaway scenarios. It is a low-speed driving different from passenger cars. It can solve problems such as delivery time, delivery demand and customer privacy and security. However, it involves a large number of internal roads and non-motorized vehicle lanes during driving, and still needs to solve external interference and other factors. Currently, companies including JD.com and Suning have begun to participate in the testing and mass production of unmanned delivery vehicles.

➢Summary: Taking a semi-enclosed hub as an example, the implementation of autonomous driving in this fixed area and low-speed environment is more direct than the various V2X problems faced by passenger cars on complex urban roads. The difference lies in which company provides a more comprehensive algorithm, a suitable mass production cost of the solution, and meets the needs of the OEM in mining areas, logistics parks and ports in the same scenario.

2.3.3 Robotaxi autonomous driving taxi service: multiple giants participate, directly enter the L4 level

➢Robotaxi is an autonomous driving taxi service, where users can book an unmanned vehicle to take them to their destination. Robotaxi services have a huge impact on travel services, so they have also attracted giants such as Google and Uber to deploy. In the existing platform and enterprise cooperation, Waymo has deployed 10 self-driving taxi service vehicles on Lyft, and Uber also plans to launch vehicle deployment operations with Toyota Motor for autonomous driving taxi services in 2021.

➢The level of autonomous driving required by Robotaxi is L4, which has very high requirements for R&D strength and safety performance. It can be adjusted according to different technical solutions in a limited area, time, and weather. Regardless of the equipment accuracy of the vehicle configuration or the system composition required by the whole vehicle, it requires high procurement costs. Although operating taxis can obtain operating income through relevant subsidies and gradually reduce the cost per kilometer, the cost of each vehicle for early test vehicles is high, and algorithm-based startups cannot cover the cost in the early stage, so most participants are industry giants.

➢The impact of Robotaxi on travel services is shown in the example of online car-hailing and private cars. After reaching a certain penetration rate, autonomous driving car-hailing services can improve efficiency through systematic coordination, and are stronger than traditional online car-hailing services in terms of cost, efficiency and safety.

➢Currently, online car-hailing services have been implemented relatively quickly in Arizona, the United States. This is partly due to Waymo's rapid implementation and continuous research and development capabilities in L4-level Robotaxi. In China, due to the government's open attitude towards the autonomous driving industry, road conditions, and supporting hardware facilities, we believe that we need to wait for other scenarios to be implemented before we can launch services in this scenario.

2.3.4 Passenger car mass production: ADAS penetration varies greatly depending on the model, and L3 is difficult to achieve

➢The passenger car market has huge potential and is one of the scenarios that start-ups are vying for in the hottest stage of the autonomous driving industry. Companies including Pony.ai, uisee, and Holomatic have all received large amounts of capital financing in the early stages.

➢ In the mass production of passenger cars, whether the technology is mature enough will determine the safety of pedestrians and drivers on urban roads. Therefore, so far, both startups and OEMs are following the gradual process from L2 to L5, based on the powerful system, modular capabilities and automotive-grade products of the existing ADAS, to further research and development and trials.

➢According to data from Guosheng Securities, OEMs including Cadillac, Geely, Great Wall, Changan, and SAIC have all launched their own L2-level models and are exploring the possibility of cooperation with algorithm-based companies. Already implemented cooperation includes the L3-level cooperation between Great Wall and AutoBrain.

➢ At present, the penetration rate of ADAS in different models is very different, and some ADAS functions are also very different. According to the data of all fuel vehicles crawled by Guosheng Securities (2,872 models ), the overall penetration rate of cruise control, blind spot monitoring, lane keeping, collision warning, and adaptive cruise control is relatively low in low-priced models, reaching 57.1%, 17%, 17.2%, 19.3%, and 17.2% respectively. The penetration rate of high-end parking spaces is relatively high, and the penetration rate of different auxiliary systems in models above 400,000 is above 50%. In the future, the price of ADAS hardware will drop as domestic hardware is replaced, and the whole vehicle will penetrate into models below 300,000 yuan.

➢If L3 level vehicles are to be put into operation, only by carrying out mass production at the same time as pre-installation can the procurement cost (including LiDAR, millimeter wave radar, etc.) be reduced through the scale effect.

➢Currently, the unit price of automotive-grade products from domestic and foreign equipment providers is relatively high, especially for chips and sensors. Algorithm-based companies conduct vehicle testing and trial runs through single small-batch purchases. Therefore, whether or not to place batch orders with the OEM is a key factor in testing whether an algorithm-based company has core competitiveness.

3. Technology industry chain and competition pattern: Sensor solutions vary greatly, and algorithms are the core

➢ The penetration of the autonomous driving industry in different scenarios cannot be separated from the support of algorithms and hardware equipment. In different scenarios, the sensor equipment, chips, and computing methods used by various companies are the technical advantages that distinguish the competitiveness of the companies.

➢ For different levels of autonomous driving, mature (L2 level) equipment support is relatively complete, including lidar, camera vision solutions, chips (such as Mobileye's EyeQ series products), etc., which already have mature first-tier component suppliers.

➢For the algorithms and systems that are gradually being developed (L3 level and above), the requirements for hardware and algorithms are becoming higher and higher, and the automotive-grade standards are gradually improving. This has given some startups the opportunity to explore business opportunities. How to provide competitive equipment in the new R&D stage has become the focus of efforts.

3.1 Industrial Chain: Perception-Decision-Control

➢From L3 level onwards, system operation will require deep learning of data collected by sensors, simulating situations that may occur in actual scenarios and conducting repeated drills. The standards of each link involved in the industrial chain will also be continuously improved.

➢Competition at the sensor level is relatively more intense. Many manufacturers are actively creating radar and camera solutions that meet the specifications , while algorithm companies are also strengthening computing power and accelerating implementation based on their own models and target scenarios.

3.2 Segmented Industry Structure: Current Status of Sensors, High-precision Maps, Chips, and Controllers

➢With the pattern of subdividing the industry, we mainly focus on the perception layer and decision-making layer, and the wire control of the control layer The installation is mainly determined by whether the OEM opens the bottom-level wire control.

3.2.1 High-precision map concept and order implementation

1) Difference from traditional map vendors: After collecting data, it is matched and converted with existing map data

➢ Traditional map providers: When a driver is driving a car in a city or on a low/highway road, the navigation map will recommend one or several routes to us. Most existing navigation maps even show congestion conditions and the time required for each route. After obtaining this information, the driver decides whether to go straight or turn based on the information provided by the map, and evaluates the surrounding driving environment. He may also consider traffic control: signal lights, speed limit signs, etc.

➢ High-precision maps: Autonomous driving cannot independently determine its current location and identify traffic lights, signs, pedestrians and other obstacles based on GPS without human intervention. Therefore, high-precision maps containing a large amount of driving assistance information have become an indispensable part of autonomous driving. High-precision maps have many characteristics, with an accuracy of 5 to 10 centimeters (high precision), (inclusiveness), and more semantic information (real-time).

➢The production process of high-precision maps is complicated. The collection vehicle collects laser point cloud and other data collected by sensors through crowdsourcing, further cleans the data and reduces noise and desensitizes it, then imports the data for job allocation, and finally merges and compiles the map to reconstruct the three-dimensional scene.

2) Business model: Fees are paid based on annual service volume

➢The production and application of high-precision maps need to be carried out through cloud services. Therefore, compared with the traditional map license model, the payment method based on year (annual fee) or service volume (service fee) is a special feature of high-precision maps. Even overseas map vendors such as Here and TomTom have not yet formed a relatively specific and clear business model for their high-precision map products. The final charging model for high-precision map products needs to be jointly negotiated and determined by major participants in the industry chain, such as map vendors, Tier 1, and car manufacturers.

3.3.2 Competition among LiDAR, MMW Radar and Camera Sensors

1) Difficulty and price of automotive-grade Lidar equipment

➢LiDAR uses 3D modeling to build data models through point cloud data. In the autonomous driving industry, the data obtained by LiDAR is combined with high-precision maps, and then returned to the actual scene for decision-making with the help of deep learning algorithms.

➢Automotive-grade LiDAR currently has 1-line, 4-line, 8-line, 16-line, 32-line and 64-line. The higher the line bundle, the higher the response speed and accuracy, but the higher the cost. Currently, 16-line LiDAR is more widely used in low-speed scenarios and parks.

➢The main participant in the automotive-grade LiDAR market is Velodyne, which sells 64-line, 32-line, and 16-line products, with official prices of US$80,000 (about RMB 523,000), US$40,000 (about RMB 260,000), and US$8,000 (about RMB 52,000 ). The high price of the products is due to the strong supply and demand relationship and Velodyne's strong R&D capabilities in the process of algorithm-based companies developing .

➢ Domestic LiDAR chips: Domestic companies involved in the development of LiDAR include Beike Tianhui, Digital Green Earth, and Leishen Intelligent. Beike Tianhui currently has self-developed products including LiDAR chips, and Leishen Intelligent has also released multi-line and solid-state LiDAR chips.

➢Solid-state LiDAR: MEMS (micro-electromechanical system), OPA (phased array) and Flash are the three main technical paths for vehicle-mounted solid-state LiDAR. Compared with mechanical LiDAR , solid-state LiDAR can better meet the requirements for the popularization of autonomous driving: large-scale, low-cost, and automotive grade.

2) Millimeter-wave radar: 77GHz and 24GHz millimeter-wave radars have more obvious advantages

➢Millimeter wave usually refers to the frequency domain of 30 to 300 GHz, which can distinguish and identify very small targets and multiple targets at the same time. Compared with ultrasonic radar, millimeter wave radar has the characteristics of small size, light weight and high spatial resolution. It has a more mature market and technology, and its anti-interference ability is better than other vehicle-mounted sensors. Due to the improvement of technology and large-scale application in ADAS, the price is relatively more reasonable.

➢ The mainstream automotive millimeter wave radars in the market can be divided into two types according to their frequencies : 24GHz millimeter wave radar and 77GHz millimeter wave radar. Generally, the detection range of 24GHz radar is short to medium distance, and it is used to realize blind spot detection system, while 77GHz long-range radar is used to realize adaptive cruise system.

➢The market share of millimeter-wave radar is basically monopolized by foreign Tier 1 companies, with manufacturers represented by Bose, Continental, Hella and Delphi monopolizing the vast majority of the market share. Among domestic listed companies, Desay SV has mass-produced 24GHz millimeter-wave radars, and among non-listed companies, Xingyidao Technology, Muniu Technology, Suzhou Millimeter Wave and other companies have received aftermarket orders.

3.3.3 Chips: ASIC chips are more suitable for automotive standards, while GPUs are widely used.

➢ Due to insufficient computing power, traditional CPUs are unable to meet the needs of processing unstructured data such as videos and images . However, GPUs can handle a large number of simple computing tasks at the same time, replacing CPUs as the mainstream solution in the field of autonomous driving.

➢In the process of ADAS evolving towards a higher level of autonomous driving, data including LiDAR point clouds and computer vision images need to be received, analyzed, and processed . Therefore, algorithm-based companies have great differences in their demand for and types of chips.

➢Among the current mainstream chip types, GPU is good at cloud training, but its power consumption is relatively high and its reasoning efficiency is average; FPGA chips have strong computing power but also high power consumption. After each burning, they will have the functions of the new connection method; ASIC chips are special chips specially designed for specific purposes of specific customers. Their low power consumption and small size meet the standards of automotive-grade products.

➢Mobileye, which has a large say in assisted driving chips, currently has the largest shipment volume. Its fourth-generation products, mainly EyeQ, can process data from 8 cameras. Mobileye has been acquired by Intel. In addition, Nvidia's GPU chips are also currently being used on a large scale. In the early R&D and trial stages, although the power consumption and cost are high, the flexibility is relatively strong.

➢In addition, chip companies including NavInfo and Horizon Robotics have also developed automotive-grade chips and computing platforms to help chips better assist autonomous driving systems.

3.3 Competitiveness: Startups must have order delivery capabilities, and large automakers focus on acquisitions

➢Here we discuss enterprises of different natures and their respective technology layout ideas.

➢Large companies - technology-driven: Internet giants and technology companies, mainly Alphabet and Baidu have solutions and algorithms, purchase cars from OEMs for modification, and build fleets to provide autonomous driving travel services. Considering different scenarios, the technology varies greatly. For example, companies focusing on Robotaxi scenarios, such as Waymo, will purchase two types of radars, solid-state and mechanical, for comparison, and will also develop their own research and development. At the same time, considering the high performance required by the operating market, the types of hardware will be richer and the accuracy will be higher.

➢ OEMs - orders, self-development or acquisitions:

➢1. Orders: The main cooperation mode of OEMs is to cooperate with technology companies. On the one hand, during the early R&D of technology companies, OEMs are not inclined to develop the underlying confidential wire control content for reference by technology companies. Technology companies generally purchase vehicles for R&D and sensor combination.

➢2. Self-research: Large car manufacturers also tend to establish their position through their own research and development in different scenarios, but the speed of research and development is slower than that of technology companies, and the implementation time is further delayed.

➢3. Acquisition: Acquiring core technology vendors is a method that large OEMs have adopted when their own research progress is delayed. Representative examples include Intel's acquisition of Mobileye, a core algorithm and chip developer, and General Motors' acquisition of Cruise, both of which were "one-step" actions taken when the research and development progress was relatively slow.

➢ Travel service providers - self-developed algorithms: The leading travel service providers will vigorously develop algorithms, purchase vehicles from OEMs, or promote mass-produced high-level autonomous driving vehicles to enter the platform. However, compared with the former , the latter has weaker bargaining power and is easily replaced by competitors.

➢ Technology companies - self-developed algorithms, cooperation with OEMs: The research and development direction of technology companies is centered on algorithms. Most of them are equipped with cameras as one of the visual devices, and most of them are equipped with laser radars for perception. The cooperation orders with OEMs are the guarantee of hematopoietic ability.

4.1 Autonomous Driving Industry Chain Map

➢ Alphabet, the parent company of Google, to which Waymo belongs, is one of the earliest pioneers in the world to invest in the autonomous driving industry. According to the driverless car test report released by the California Department of Motor Vehicles (DMV) in 2017, the total test mileage of the 75 vehicles invested by Waymo at that time reached 566,000 kilometers, and the average driving distance was 5,596 kilometers before the safety officer needed to take over manually.

➢ Automated ride-hailing service: Waymo's service has been launched in four areas in the suburbs of Phoenix, Arizona, USA. Currently, Waymo is cooperating with FCA and Jaguar Land Rover on autonomous ride-hailing services, purchasing vehicles from these two OEMs for transformation and operation. At the same time, it has also reached a cooperation with Lyft on online ride-hailing services, temporarily deploying 10 driverless cars on the Lyft ride-hailing platform , with service prices similar to those of Uber and Lyft.

➢ Waymo cooperates with top Tier 1 Magna to build a factory to produce L4-level vehicles. The factory is located in Michigan, which has advanced automotive high-tech technology and a mature manufacturing system . Such a transformation plant is different from a traditional manufacturing plant. It is a secondary transformation of the whole vehicle by Waymo and invested in the Robotaxi business. The reason is the confidentiality of its core competitiveness (including core hardware such as sensors and central computing platforms).

➢ Morgan Stanley's valuation of Waymo at $175 billion is based on its three business scenarios: autonomous driving ride-hailing services, logistics and transportation, and technology licensing to vehicle manufacturers. The estimated charging standard for autonomous driving ride-hailing services is $0.9/km, which, combined with a 4% global market share excluding China, is valued at about $80 billion; the "last mile" transportation service for home delivery is about $90 billion; vehicle manufacturers pay about $7 billion in technology licensing fees to obtain Waymo's autonomous driving technology and algorithms.

➢ NVIDIA's investment and focus on autonomous driving are different from Mobileye's, and the chip specifications and features are quite different. In Chapter 3, we analyzed that in the long run, ASIC will be the mainstream of mass-produced automotive-grade chips required for autonomous driving above L3 level. Mobileye, the industry leader of ASIC, has a high actual application volume and market share, but its attitude towards technology openness was unclear in the early stage, and algorithm-based companies also encountered obstacles in further improvement and adjustment.

➢ In 2015, NVIDIA released DRIVE PX, which is used for environmental information recognition and processing in self-driving cars. Based on the general GPU architecture, it opened up the AI ​​platform NVIDIA DRIVE to enable OEMs, Tier 1 and start-ups to accelerate the production of self-driving cars.

➢ NVIDIADRIVE can fuse data from multiple sensors to perceive the surrounding environment in all directions, and use deep neural networks to improve the accuracy of data fusion.

➢ From the existing mass production situation, DrivePX2 equipped with GPU has been put into mass production, with a unit price of 16,000 US dollars and a power consumption of 425 watts. It has been installed on Tesla's Model S and Model X. Xavier is an evolutionary version with significantly improved performance and a power consumption of only 30 watts.

➢Business idea: Based on the research and development of subsystem modules such as perception, sensor fusion, high-precision maps and simulation, using its own general technology and matching different scenarios, the perfect products obtained in the process are put into operation, providing complete system and algorithm solutions to OEMs and operators , and applying standardized products to toB and toC directions respectively.

➢ Landing scenarios: involving pre-installation of passenger cars, pre-installation of commercial vehicles (trucks/vans), and post-installation (renovation of old vehicles in the park). Among them, in terms of passenger cars, in the pre-installation link of new car mass production, AutoBrain's cooperation idea is to focus on selling the completed solutions to the OEMs; in terms of the landing and operation capabilities of commercial vehicles, some post-installation links are also involved, but AutoBrain is more inclined to the pre-installation link, including the shuttle bus in the park and the unmanned driving realization capability of logistics transportation in the last mile.

➢ AutoBrain's key businesses are fleet and logistics, including L3 passenger cars, L3 trunk logistics, L4 branch logistics and L4 last mile.

➢ In the construction of commercial scenarios, cooperation with local governments and operators is the main focus. In the cases of cooperation , taking the passenger car Great Wall WEY as an example, the car is equipped with 6 millimeter-wave radars and 2 cameras, which can realize autonomous overtaking, autonomous lane change, autonomous on-ramp and off-ramp, etc., and has reached the L3 level vehicle standard.

➢Advantages of commercial vehicles: AutoBrain has a relatively mature model of cooperation with OEMs in the domestic logistics and transportation scenarios. In the process of negotiating with operating companies and cargo owners, it has a complete transportation capacity solution. The operation solutions are mainly divided into two categories: single vehicle and formation.

✓ Single vehicle: The single vehicle mode serves the scenario from highway toll station to toll station. This operation plan saves 33-50% of labor costs, and the short-term benefits are relatively fast. In the long run, the advance planning of the route can also save fuel consumption for vehicle acceleration, deceleration and slope control, saving 13% of the current fuel consumption.

✓ Formation: The platoon operation plan includes a lead vehicle and four subsequent L4-level vehicles, which can help save 90% of driver costs and 20% of relative fuel consumption costs.

✓The charging model currently uses a bicycle as a unit, providing the AutoBrainL3/L4 level solution including hardware, software, domain controller, etc., and charging according to different operating scenarios.

➢ Unmanned bus business: AutoBrain also cooperates with local governments and operators to pre-install unmanned bus services on some buses in Beijing and Tianjin. Among them, the single-vehicle unmanned bus operation scenario is further subdivided into fixed routes and non-fixed routes. Fixed routes include high-speed BRT routes with relatively simple road structures; non-fixed routes generally select random point-to-point routes (A to B) with average road complexity.

➢Technical path: At the technical level, AutoBrain conducts L4-level research and development of different module systems, and uses deep learning algorithms to identify, track, and make more accurate predictions. AutoBrain is based on multi-modal sensor fusion, uses machine learning to make decisions, and predicts control through MPC models, and puts the entire software system into a small automotive-grade domain controller.

➢AutoBrain's current negotiations with OEMs are at different stages. For OEMs with a high penetration rate of autonomous driving background, the negotiation pace is relatively fast. Due to their mastery of L3-level autonomous lane change and other technologies, their bargaining power is also relatively high.

➢For OEMs and Tier 1s who have a relatively low understanding of the technical difficulty of the autonomous driving industry, AutoBrain will negotiate and communicate with them based on the idea of ​​"POC (proof of concept) - POV (proof of value) - SOP (mass production)". Although this process takes a relatively long time, it is a long-term understanding process for both customers and companies, which is more conducive to the subsequent cooperation.

➢ TuSimple was founded in September 2015, focusing on L4 driverless truck solutions, with R&D centers in both China and the United States. TuSimple implements autonomous driving in two scenarios: trunk logistics scenario and semi-enclosed hub scenario.

➢ In the early stages of research and development, TuSimple has been committed to the L4 level of commercial trucks because of the cost of trucks and labor costs in logistics scenarios. In commercial scenarios, due to the high operating costs and the shortage of truck drivers, L3 level autonomous driving still cannot solve these two problems. Therefore, by directly addressing the pain points, L4 level autonomous driving can be achieved through system autonomous control.

➢ Trunk logistics high-speed freight scenario: In the trunk logistics scenario, considering the differences between China and the United States, the implementation time is different mainly due to the relevant supporting policies and road approval. TuSimple has already started warehouse-to-warehouse paid commercial transportation in Arizona and Texas, USA, and the business model is based on a per-kilometer payment model based on the transportation distance.

➢ The pain points of trunk freight can be generally divided into four points: operating costs, traffic accidents, driver shortages and environmental issues.

➢Pain points: From the pain points of each scenario analyzed in Chapter 2, driver costs, truck mileage, and fuel consumption are several important influencing factors.

➢According to statistics from the American Trucking Association and the United States Department of Transportation, there are more than 415,000 truck-related traffic accidents each year, accounting for 30% of all automobile traffic accidents. There is a huge shortage of drivers, and the work intensity is high, making the profession relatively unattractive. TuSimple's technology application solutions also focus on these issues.

➢ Technology: TuSimple uses 1,000-meter visual perception technology at the perception layer to enable trucks to make advance predictions to the greatest extent possible. It combines 3D cameras, two laser radars, IMU and other sensors, and NVIDIA's GPU chips to plan the optimal driving route in structured high-speed scenes and unstructured low-speed scenes, solving traffic accidents caused by emergencies and line of sight problems , and reducing repeated driving caused by unfamiliar roads. In addition, TuSimple's night camera perception system can realize the all-weather operation of unmanned trucks, helping truck operating companies improve efficiency and reduce costs, and expand transportation capacity.

➢ Closed/semi-closed park scenarios: In this type of scenario, since most of the technology can be shared, combined with TuSimple's autonomous driving technology solution, the R&D costs can be spread across different scenarios. As for the implementation of park scenarios, since the actual application of pure unmanned ports can only be used for newly built ports, the direct transformation of old ports will not be invested in large-scale transformation due to factors such as high cost and long trial operation cycle. Therefore, some after-sales wire control transformation is carried out in combination with existing trucks, and trial operation vehicles developed by TuSimple are introduced for matching.

➢ In the future, based on the application scenarios of existing technologies and the business models that have been implemented, TuSimple will explore more cooperation models to meet the business needs of different customers.

5. Summary and trend forecast: The implementation time of L3 level continues to be postponed, and the Internet of Vehicles will bring changes

➢At present, the national policy has issued a series of autonomous driving road test details and management methods, including the "Intelligent Connected Vehicle Road Test Management Specifications (Trial)", to help the autonomous driving industry accelerate its implementation. Various provinces and cities have also introduced relevant methods for autonomous driving road testing to promote commercialization. In China, companies including Internet companies and technology companies, OEMs, and travel platforms have 109 autonomous driving road test licenses, of which Beijing has the most, with 59. The companies involved are rich in variety, including OEMs, Internet companies, travel platforms, etc. Chongqing ranks second with 12 licenses, and the participants are mainly OEMs.

➢Despite this, there are certain risks in going on the road for different scenarios of passenger cars and commercial vehicles. Issues including laws and regulations, consumer acceptance, safety, etc., make all parties dare not be too open. However, with the gradual maturity of the L2 level in the OEM, the use of ADAS functions in mid-to-high-priced models has become wider, and gradually penetrated into mid-to-low-end models. Coupled with the technological breakthroughs of technology companies in the L3 level passenger car scenario, the time for the OEM to mass-produce the L3 level may be after the technology reaches the standard, and then it can be implemented after a large number of road test simulations and trial runs. Commercial vehicles need to rely more on scenarios to make judgments. In areas such as Arizona in the United States where independent legislation can be made, there are large trial operation areas, which are convenient for technology companies to simulate and try and error, and further explore the pre-installation of heavy trucks. The closed park scenario relies more on post-installation modification of wire control.

➢We have already discussed and analyzed the technology, scenario feasibility, and key company cases. Here we will make a simple analysis of the industries related to autonomous driving.

➢ Intelligent connected vehicles, as core products in important strategic documents, involve industries such as new energy, autonomous driving, and Internet of Vehicles. For the autonomous driving industry, electronic core components including sensors, processors, chips, etc. involved in mass production account for a higher proportion of the vehicle cost. Compared with the existing automotive products with a life cycle of 3 to 5 years, technology companies can better understand automotive wire control and hardware equipment than before, so they can leave room for subsequent L3 and L4 upgrades. As the accuracy of V2X equipment in the Internet of Vehicles increases, it will make more timely judgments on high-density roads and complex road conditions, helping to speed up the key links of autonomous driving.

Since mining belongs to the edge application field, AI is still the main application field of ASIC. With the advent of the artificial intelligence era, traditional neural network algorithms are not efficient on general-purpose chips (CPU, GPU) and consume a lot of power. Therefore, from the perspective of chip design, general-purpose often means higher costs.

In order to improve efficiency and reduce power consumption, ASIC came into being. At present, from a global perspective, the field of ASIC based on artificial intelligence has not seen a "monopoly" situation, but instead has presented a pattern of competition between domestic and foreign electronic technology giants, research institutes and domestic start-ups. Foreign companies are led by Google, IBM, Intel, and Stanford University, while domestic companies include Vimicro, Cambrian Technology, and Qiying Tailun.