Emerging technologies are often the most difficult to "land". Many people define 2021 as the first year of autonomous driving, at this stage, the trend of electrification and intelligence is impacting the traditional automobile industry chain, and the epidemic has made domestic users have a strong unmanned demand. However, Qi Ji can't take ten steps, and it still has a long way to go to achieve full automation of L4-L5.
Author 丨 Xiaoyu
Edit | Sibe
Reviewing the autonomous driving technology since 2021, whether it is AI chips or vehicle-level sensors, there are multiple milestones. At the same time, technology has spawned intelligent scenarios, multi-industry concerted efforts, in the epidemic, trade war, lack of core tide and other black swan events catalyzed, autonomous driving terminal applications burst out of unprecedented demand, but the birth of new technologies is always accompanied by large or small problems. In 2022, autonomous driving, how many "hurdles" will be taken?
1
It is difficult to increase the computing power of automation
On March 1, 2022, the National Standard for Automobile Driving Automation Classification issued by the State Administration for Market Regulation for autonomous driving functions will be officially implemented.
With the improvement of the level of automatic driving, the required computing power increases at a high speed. The level of intelligence of autopilot depends on whether the algorithm is powerful, from L1 to L5, every level of autonomous driving is increased, and the computing power requirement is also increased by one level.
At present, driver assistance technology has been deployed in production vehicles, known as "advanced driver assistance systems". The world's current autonomous driving assist technology is still only between the L2-L3 level.
Before L3, the computing power required for automatic driving was low; the AI computing power required by L3 reached 20TOPS; after L3, the computing power requirement increased dozens of times, L4 was close to 400TOPS, and L5 computing power requirements were more stringent, reaching 4000+TOPS. For each additional level of autonomous driving, the demand for computing power increases by an order of magnitude. According to Intel's calculations, in the era of fully autonomous driving, the amount of data generated by each car per day will be as high as 4,000 GB.
2
AI chips are difficult to get on the car
With the increasing level of intelligence of autonomous vehicles, the amount of data that needs to be processed is getting larger and larger, and software and hardware devices such as high-precision maps, sensors, and lidar put forward higher requirements for computing, so in addition to high-precision sensors, AI acceleration chips have become the mainstream.
To evaluate the performance of the chip, computing power, energy consumption, and efficiency are indispensable.
At present, there are only a few automotive-grade AI chips available on the market. Foreign companies such as Tesla, NVIDIA, Intel, and Moblieye are the mainstays, while domestic Startups such as Huawei HiSilicon, Cambrian, and Horizon are "taking advantage of the momentum."
In 2021, car companies began to tilt towardS NVIDA in the choice of automatic driving chips, and also chose large computing chips such as NVIDA Orin X, Qualcomm Snapdragon Ride, and Huawei MDC in the layout of future models, and more and more OEMs adopted high-computing automatic driving chips.
It can be seen that the AI chip wants to "get on the car", and the core is "high speed, high precision, and high accuracy".
Taking the Audi A8 as an example, the whole car has a total of 12 ultrasonic sensors, 4 panoramic cameras, 1 front camera, 4 medium-range radar, and 1 infrared camera. In addition, there are ultrasonic radar, millimeter wave radar and lidar. These sensors work simultaneously and can generate several gigabytes of ambient detection data per second.
Below 60km/h, the autopilot system can perform actions including starting, accelerating, steering, and braking, and once the autopilot system reaches its limit, the driver is immediately notified to regain control of the driving. This is also considered a breakthrough from L2 to L3 after Tesla Autopilot 2.0.
In addition to chips and sensors, it is necessary to mention high-precision maps.
Self-driving cars need to know exactly where they are on the map, including data on the slope, curvature, heading, elevation, and roll of each lane. The types and colors of the lane lines; the speed limit requirements and recommended speeds of each lane; the width and material of the barrier belt; the arrows on the road, the content of the text, the location; the absolute geographical coordinates, physical dimensions and characteristics of traffic participants such as traffic lights and crosswalks... All of this information also needs to be accurately reflected in high-precision maps.
Autonomous vehicles at L3 and above must rely on their precise positioning to effectively bridge sensor performance boundaries. For autonomous driving, high-precision maps are an important guarantee for achieving highly automated driving.
At present, Baidu, NavInfo and AutoNavi occupy the main shares, and the domestic market presents a "three-legged stand". According to IDC statistics, the top five companies in the domestic high-precision map industry market share in 2020 are Baidu, NavInfo, AutoNavi, Yitutong and Here, of which CR3 exceeds 65%, showing a "three-legged" situation.
3
The cost of lidar remains high
At present, there are two major perception paths in the mainstream market of automatic driving, "weak perception + super intelligence" and "strong perception + strong intelligence", and Tesla uses the former.
"Weak perception + super intelligence" refers to the main reliance on cameras and deep learning technology to achieve environmental perception, rather than relying on lidar. The image transmitted back by the camera is calculated by the algorithm, and then the operation is achieved by the controller and the actuator, which is Musk's option, but the so-called "super intelligence" requires a strong enough terminal arithmetic and algorithm to provide support.
On Feb. 10, U.S. regulators said Tesla was recalling 578607 cars in the U.S. because pedestrians might not be able to hear the necessary sirens that sounded as cars approached.
In addition, Tesla still has certain problems in different scenarios, such as driving into the tram track, stopping at a distance, not being able to recognize some traffic signs, and hitting a baffle. In order to achieve high-precision environmental perception, there are more detailed technical difficulties that need to be studied and solved.
Different from Tesla's persistence in "camera + artificial intelligence", many domestic autonomous driving top forces have bet on the sensor on lidar and taken a road of "strong perception + strong intelligence". Google Waymo, Baidu Apollo, Uber, Ford Motor, General Motors and other self-driving companies, as well as traditional car companies, are in the "strong perception + strong intelligence" technology camp.
Among them, Baidu chose Hesai, including relatively mature products such as mechanical rotary type (Pandar series) and MEMS micro-galvanometer (GT series) based on dTOF; Huawei's lidar technology route chose the semi-solid MEMS micro-galvanometer technology route.
At the beginning of 2022, NIO is expected to release a new model such as ET7 and ET5 with an ultra-long-range high-precision lidar, as well as seven 8-megapixel high-definition cameras, four 3-megapixel high-sensitivity surround view cameras, 5 millimeter-wave radars, 12 ultrasonic sensors, 2 high-precision positioning units, and V2X vehicle-to-road collaboration.
At present, the price of mechanical lidar is very expensive, and the official price of Velodyne's 64-line/32-line/16-line products on sale is 80,000/40,000/8,000 US dollars, respectively.
It is estimated that from 2022 to 2025, the demand for lidar in the global passenger car market will increase from 220,000 to 21.34 million, and the penetration rate will increase from 0.2% in 2021 to 14% in 2025, when the cost is expected to decline.
4
Policies and regulations are difficult to implement
Technology landed, safety first, automatic driving is still a certain distance from the real "maturity". Legislation becomes complicated when highly ordered procedural control is combined with highly disordered manned driving. Who should give control of the steering wheel has become a problem in the field of autonomous driving.
For the acceptance of new technologies, China often adopts a more stable strategy.
In 2021, the Cyber Security Bureau of the Ministry of Industry and Information Technology and the Quality Development Bureau of the State Administration for Market Regulation issued a number of policies related to automotive network security, but the mainland has not yet issued a special legal norm for autonomous driving data, and its related provisions are scattered in other laws.
At present, there is still a lack of special laws and regulations on autonomous driving data in China, the data ownership right of automobiles when collecting geographic information is still unclear, and the specific definition of "important data" has not yet been clarified.
Compared with China, in December 2021, Germany legally recognized L3 autonomous driving on the road.
At the same time, Mercedes-Benz became the first car company in Europe to meet the UN-R157 regulations, becoming the first brand in the world to use L3 assisted driving on open roads, but this was also limited to the use of this feature on the highways in Germany.
According to SAE's definition of L3 level automatic driving, when L3 automatic driving is turned on, if there is an accident, the driver is not responsible, but the car company is responsible for it.
In order to avoid various unexpected situations, Mercedes-Benz still made strict regulations on the opening conditions of L3 automatic driving, the system can only be opened at a speed of less than 60km/h, although the system allows the driver to leave the steering wheel with both hands, but still keep both eyes to observe the road surface conditions in order to take over the vehicle at any time.
At present, the parties to the UN-R157 regulation are EU countries, the United Kingdom, Japan, South Korea, Australia, etc., and Mercedes-Benz's L3 level autonomous vehicles can be sold in these countries.
Emerging technologies are often the most difficult to "land". Many people define 2021 as the first year of autonomous driving, at this stage, the trend of electrification and intelligence is impacting the traditional automobile industry chain, and the epidemic has made domestic users have a strong unmanned demand.
However, Qi Ji can not take ten steps, in order to achieve the full automation of L4-L5, automatic driving still has a long way to go, and there are many technologies that are not optimistic at the beginning. The conditions for commercial landing are mature enough, and it is hoped that autonomous driving will move towards a more open scene.