(By Caixin journalists An Limin and Bonnie Cao)
After years of debates and trials of different technological paths to autonomous driving, Chinese smart-car makers have gradually steered toward the methods proven by Tesla Inc. and are gearing up for a race against the industry giant.
By 2024, the performance of Chinese autonomous smart cars will match that of Tesla at the end of 2022 and early 2023, said Li Auto CEO Li Xiang.
The Society of Automotive Engineers (SAE) defines six levels of driving automation ranging from 0 for fully manual to 5 for fully autonomous. Global automakers are now racing to reach Level 4, which enables vehicles to operate without human intervention under most circumstances.
The commercial prospects for autonomous vehicles are promising, but the road leading to them is bumpy as automakers are still in search of effective, affordable technology solutions.
“To many customers, assisted driving features may not be as attractive as leather seats,” said Ai Rui, vice-president at Haomo.AI Technology, a Beijing-based autonomous driving startup, referring the current limits of the technology.
The next step toward full autonomy will be to enable automated driving on urban streets in the presence of a driver...
The next step toward full autonomy will be to enable automated driving on urban streets in the presence of a driver, which may become a turning point for consumers to directly understand autonomous technology, industry analysts said.
Driving on city streets with an urban assistance system will provide a completely different experience for consumers, Li said in March.
Tesla is a pioneer in autonomous driving with its “Autopilot” mode. It started in 2020 with the development of a Level 2 system that works on city streets. Tesla said about 400,000 car owners in the US and Canada had access to test versions of the system as of the end of 2022.
Chinese companies have placed their bets on lidar technology and high-precision maps to detect and forecast road conditions...
Most autonomous driving systems deploy a full suite of sensors that include cameras, lidar — laser-based sensors — and radar. Chinese companies have placed their bets on lidar technology and high-precision maps to detect and forecast road conditions. In 2020, XPeng said it would implement lidar technology in its autonomous vehicles, and it later acquired a high-precision mapping
In contrast, Tesla has pursued a vision-only approach based on cameras and neural network processing. It achieved positive results while other companies have been struggling to show value in their assisted driving capabilities. Lidar and high-precision maps require high investment, making them increasingly difficult to develop amid a slowing market.
Since 2023, carmakers have fallen into a fierce price war as supply has outpaced demand. After Tesla cut prices earlier this year, most autonomous vehicle companies quickly followed suit.
XPeng said it will deploy its automotive driver assistance systems for urban roads without high-definition maps across China during the second half of 2023. Li Auto said at the Shanghai International Automobile Industry Exhibition last week that its latest autonomous driving model will be able to function without high-precision maps. Huawei Technologies Co. Ltd. and Horizon Robotics, a Chinese chipmaker and partner of Volkswagen AG, both signalled last month that they will release similar products.
It’s costly in terms of time and money to develop high-precision maps. They require scanning and mapping by specialised vehicles equipped with various sensors, a technician from an autonomous car startup told Caixin. The average cost is 1,000 RMB (US$144.37) per kilometre, and the maps need continuous maintenance to keep up with actual changes of road conditions.
“We have spent so much time, and we haven’t even finished mapping all the small streets in Shanghai,” said Richard Yu, consumer and auto division head of Huawei. He observed that it would be prohibitive to make nationwide maps.
Legal and regulatory requirements also hinder the development of high-precision maps as only qualified companies can do the mapping. The regulatory approval for each update due to road changes is often long and complex, XPeng chairman He Xiaopeng said.
An increasing number of companies are shifting toward Tesla’s camera-based approach to identify and differentiate between objects. The data collected can also be used for simulation training, to test the system’s response capabilities under rare accident scenarios. It will reduce the cost of training to 1% of what it is now if companies can simulate such scenarios, according to Haomo’s Ai.
There are split views over lidar. Li Auto’s high-end car models are equipped with lidar sensors from Hesai Group, while Nio Inc. uses devices made by Suzhou-based Innovusion Holdings Ltd.
Car companies use lidar as a selling point, just as new cell phones and computers emphasise chip processor and camera performance, a product manager from a car company said. Consumers think cars with lidar sensors look “very high-tech”, according to market surveys.
Lidar sensors are both applauded and hated in the industry. They don’t require daylight to work, whereas cameras do. Lidar can calculate accurate distances to many objects that are simultaneously detected. But its high cost puts many companies off. In a car, the price of a single lidar sensor is almost equal to the sum of all other sensors combined, said Zhang Lichen, a department head of lidar startup Tanway Technology.
Lidar makers have cut prices sharply to win more orders. Chinese lidar maker Hesai slashed the price of its devices from about $17,400 in 2019 to $3,100 in 2022.
The company’s gross margin fell to 30% in the fourth quarter, from 44% in the first three quarters of last year. It posted a net loss of 300.8 million RMB in 2022.
With improvements to Tesla’s pure vision approach, more analysts are pessimistic about the outlook for lidar. One industry professional said he thought the next step in the industry’s development will be to drop lidar and move toward a purely visual approach.
It takes more than high-precision maps and lidar to match Tesla’s pure vision approach, which is not a shortcut and is more difficult.
Some still see a future for lidar. When it comes to Level 4 autonomous driving, the requirements for system safety redundancy are higher, and lidar may be indispensable. However, at the assisted driving stage, the high cost of lidar may well be a hindrance, Haomo’s Ai said.
Wu Xinzhou, XPeng’s vice-president in charge of autonomous driving, said in March that urban traffic scenarios are complex, and lidar serves as a basic safety net. As visual capabilities improve, companies and the industry are exploring whether lidar will still have a role to play in future assisted driving or high-level autonomous driving, Wu said.
It takes more than high-precision maps and lidar to match Tesla’s pure vision approach, which is not a shortcut and is more difficult. It means that the system has to process a lot of video data both on vehicle and in the cloud.
Tesla uses its self-developed chips and doesn’t use lidar, lowering the costs of its autonomous vehicles. The company has sold more than 4 million cars, making it hard for rivals to match the massive volume of data it has accumulated.
Tesla also tries to automate a lot of the labelling, meaning that the images collected are tagged with information such as vehicles, lanes, street signs and other categories. Manually tagging 500,000 to 1 million video clips would require 2,000 people working for a year, according to XPeng’s Wu.
Chinese companies have realised the importance of car-based computing power. Nvidia Corp. said on 21 March that China’s largest electric-vehicle maker, BYD Co. Ltd., will equip its next-generation Dynasty and Ocean models with the US chipmaking powerhouse’s Nvidia Drive Orin system-on-a-chip, which can power artificial intelligence cockpits and automated driving functions.
Among other Chinese carmakers, Nio installed four Drive Orin chips capable of more than 1,000 trillion operations per second (TOPS) of computing power in some of its high-end models, while Li Auto and XPeng are using two such chips with 500 TOPS.
... the actual user experience is falling short due to the poor functioning of the driver assistance system, experts say.
Consumer expectations have been boosted by expensive hardware and fancy computing power, but the actual user experience is falling short due to the poor functioning of the driver assistance system, experts say. A 2023 McKinsey survey found that electric vehicle consumers are most concerned about battery range and charging time, followed by driving performance, operating costs and other metrics. Autonomous driving functions rank ninth.
In 2022, major Chinese smart-car makers all failed to reach their annual sales targets. As the auto market weakens further in 2023, Chinese car companies are starting to learn how Tesla balances cost and functionality. It takes time to lower the cost of good-quality hardware, but Chinese companies should speed it up as much as they can, according to a veteran on autonomous driving algorithms.
Many leading Chinese companies started to develop chips on their own, Caixin learned. XPeng, Haomo.AI and Geely have announced plans to build supercomputing centres.
The roadmap of Tesla’s autonomous driving technology is not a secret, but Chinese automakers seem to have gone around in circles before figuring out the direction they are heading in, Haomo’s Ai said.
This article was first published by Caixin Global as "In Depth: Chinese Smart-Car Makers Get Into Tesla’s Lane". Caixin Global is one of the most respected sources for macroeconomic, financial and business news and information about China.
Related: Shanghai Auto Fair: Battle of the Chinese EV giants | China’s EV battery boom goes bust | Chinese automakers rev up price war in race for market dominance | Chinese EV manufacturers face tough challenge against Tesla amid subsidy cuts