The domestic autonomous driving industry is welcoming a spring.
Recently, four departments in China jointly issued a notice on the pilot work for the access and road use of intelligent connected vehicles, which formally specified the access standards for L3/L4 autonomous driving and improved relevant regulations.
This means that the country is ready to launch L3 and L4 level autonomous vehicles, and autonomous driving has taken another step towards commercialization and scaling, which can be considered a milestone event for the development of the domestic autonomous driving industry.
▲ Four departments jointly released the notice on the pilot work for the access and road use of intelligent connected vehicles
The document emphasizes the division of accident responsibilities. In the event of violations or accidents while the autonomous driving system is activated, the pilot enterprises and the pilot users must provide proof materials to the relevant departments within the specified time. If the materials are not provided as required, they will have to bear the accident responsibility.
In addition to the improvement of policies and regulations, the relevant technology for autonomous driving has also made significant progress, and the explosion of large models has provided more ideas for autonomous driving. The leading players in the industry are all adopting large model approaches in autonomous driving. Currently, Tesla’s FSD V12, which is pushed to internal employees, adopts an end-to-end large model approach.
On the other hand, to ensure the safety of autonomous driving, simulation, closed-course, and real-road testing methods are indispensable. Typically, the sequence is to first simulate, then test in closed courses, and finally on real roads, using different testing methods at different stages of autonomous driving system development and application. In terms of the number of scenario tests, simulation tests far exceed closed-course and real-road tests, while real-road tests usually far exceed closed-course tests, thus improving testing efficiency, reducing costs, and minimizing testing time. From an economic and efficiency perspective, autonomous driving simulation testing is the most effective means, as it can cover more scenarios and better meet the OTA needs throughout the lifecycle..
▲ Urban traffic digital twin simulation testing
Many companies are already laying out in this area. For example, Tencent has proposed the concept of real-time digital twins and has conducted much thinking on simulations in the autonomous driving field. They have also established collaborations with automakers such as GAC and SAIC.
With the continuous improvement of policies and related technologies, autonomous driving is developing in a very promising direction.
In this context, recently, Zhixiang/Car Zhixiang and Tencent launched the “Digital Twin Q&A” series activities, the first phase focused on autonomous driving, inviting industry experts, scholars, major manufacturers, and representatives from the autonomous driving industry chain to engage in discussions, further analyzing the important role of simulation in the process of implementing autonomous driving.
The autonomous driving industry has made some progress in recent years due to breakthroughs in related technologies and market demand, but overall, it is constrained by high technical complexity, fluctuating demand, and imperfect regulatory policies, resulting in a fluctuating development rhythm.
Zhao Xiangmo, president of Xi’an University of Architecture and Technology, professor and doctoral supervisor at Chang’an University, and head of the national key R&D program for new energy vehicles, highlighted at the conference that the development of connected autonomous driving involves comprehensive coordination among vehicles, roads, clouds, and networks, but the core issue is to first solve the problem of low-latency and high-reliability network communication. Currently, 5G cannot perfectly support the development of autonomous driving in terms of both cost and performance, especially the cost burden on companies. If the communication technology supporting autonomous driving can be well resolved in the future, the autonomous driving industry will experience explosive growth.
For the autonomous driving industry, this is a very important moment: on the one hand, regulations related to L3 and L4 have emerged in the country, encouraging the entire industry to explore more, laying a solid foundation for autonomous driving from a regulatory perspective.
On the other hand, new technologies continue to emerge in the autonomous driving field, the most typical being Tesla, which officially proposed the combination of BEV + Transformer back in 2021, starting to develop autonomous driving with a new approach.
▲ Tesla FSD Beta version page
Subsequently, many domestic players quickly followed suit. Currently, almost all players are laying out the BEV + Transformer algorithm. In 2022, Tesla further introduced the Occupancy grid network, further improving the vehicle’s 3D spatial recognition. The rapid implementation of these technologies has also accelerated the rapid development of autonomous driving, with the current hot high-speed/city NOA relying on these algorithmic updates.
To promote the mass production of autonomous driving, the importance of simulation testing is continuously increasing. By simulating the environment and vehicle conditions while driving, a closed-loop operating environment for autonomous driving is established. Through these tests, the operational status of autonomous driving algorithms can be better verified, and vulnerabilities in the algorithms can be identified and updated earlier.
At the conference, Zhu Xichan, professor at Tongji University’s School of Automotive Engineering, director of the Automotive Safety Technology Research Institute, and doctoral supervisor, mentioned that while road testing can ensure the authenticity of the data collected, achieving full scene coverage requires more scientific methods, among which simulation is paramount in the autonomous driving toolchain. It can cover as many scenarios as possible in the shortest time and with the least mileage, while also generating data.
Autonomous driving has very high safety requirements, making real-world testing an unsuitable approach; simulation testing based on the digital world is more important.
This point runs through the entire process of autonomous driving. Taking the recently hotly discussed AEB as an example, some companies require an AEB trigger every 500,000 kilometers, while others need to accumulate testing over 800,000 to 1,000,000 kilometers.
In response to this “debate,” Changan Automobile’s Chief Intelligent Driving Technology Officer Tao Ji believes that the argument over which company’s AEB is better indicates that everyone has moved beyond the stage of “having or not having” autonomous driving functions to the stage of “usable or not usable.”
However, mileage coverage has a problem: for ordinary consumers, the probability of traffic accidents is extremely low; some may never experience an accident in their lifetime, while others may have accidents from time to time. These cases cannot be covered by mileage.
Zhu Xichan believes that scene coverage is more scientific than mileage coverage, and simulation can promote the realization of scene coverage because more scenarios can be generated in simulation testing.
▲TAD Sim virtual city cloud simulation
Tencent has already made many layouts in this area, and digital twins can provide some capabilities needed for simulation.
On the other hand, the testing and verification of autonomous driving exist in long-tail scenarios. The higher the maturity of the algorithms, the lower the effectiveness of the collected data.
For some extreme scenarios, such as collisions or rollovers, it is very difficult to collect data through real-world scenarios.
Moreover, there is a sample balance issue with the data; if some sample sizes are very small, it becomes challenging to train a well-performing model with this data, and all of this can be achieved through simulation.
Tencent Group Vice President and President of Tencent Smart Transportation and Mobility, Zhong Xiangping, mentioned that digital twins can transform data into reality, presenting the real world in a digital way, providing more scenarios and sufficient samples to ensure the safety of the results.
▲Tencent’s digital twin simulation technology achieves scene style conversion (upper left, lower left – simulation scenes, upper right – real images, lower right – images after style conversion)
Furthermore, simulation can also bring many advantages in terms of cost and time. Specifically, testing with real vehicles requires a significant investment in vehicles, leading to high costs.
On the other hand, in real testing, the accumulation of data is positively correlated with the time invested; only by spending more time can more data be collected.
From an efficiency perspective, simulation can achieve the most scene coverage in the least mileage and time, truly enhancing testing efficiency.
Tencent Maps Vice President and Head of Tencent’s Digital Twin Business, Zhang Shaoyu, noted that in autonomous driving simulation testing, the combination of maps and digital twins will play an important role, providing a more realistic and accurate simulation environment to simulate real-world roads, traffic flow, traffic signals, traffic signs, etc., facilitating more comprehensive simulation tests, including functionality, performance, safety, and user experience. At the same time, using simulation testing for pre-OTA data checks of high-precision maps can reduce the costs of map data checks and regression testing time.
.Yuanrong Qixing CEO Zhou Guang also pointed out that during real vehicle testing, various real-world factors affecting the intelligent driving system must be considered, such as the volume, angle, speed, lighting, weather, and sensor status of obstacles. However, these are not easy to replicate. To accumulate more test data and replicate extreme scenarios, intelligent driving companies need to conduct testing not only in real vehicles but also in simulation systems. Digital twins can achieve high-precision restoration of traffic scenarios, realistically simulating natural environments such as rain, snow, lighting, etc., and road conditions, enhancing the reliability of simulation testing.
▲ Digital twin simulation testing on a snowy highway
In addition to digital twins, Tencent is also combining currently popular AIGC technology to develop some solutions.
Currently, simulation testing involves collecting real data to reconstruct digital scenarios, then editing to generate more scenarios. The explosion of AI technology can significantly increase the efficiency of this work.
Tencent’s Autonomous Driving General Manager Su Kuifeng discussed that using AIGC technology can create more synthetic data, especially for sensitive or high-safety areas and long-tail data, filling gaps in real data and enhancing the diversity, completeness, and balance of training and testing samples. At the same time, leveraging large language models like ChatGPT can enhance human-machine collaboration capabilities, achieving better human-machine interaction and co-driving.
▲ Nighttime traffic digital twin simulation testing
Therefore, overall, the importance of simulation testing in the implementation of autonomous driving has reached a consensus within the industry, and Tencent has deep technical accumulation in this area. As technical capabilities continue to be released and implemented, the arrival of the L3 autonomous driving era will be accelerated.
For today’s smart cars, autonomous driving is a very important link; achieving good autonomous driving is essential for better realizing the concept of smart cars.
The development model for smart cars differs from that of traditional cars. It relies heavily on data-driven approaches, requiring more data for continuous iteration and optimization.
Zhou Guang believes that massive testing data can feed back into the algorithm framework, helping intelligent driving systems solve more long-tail scenarios. Yuanrong Qixing has built a big data problem statistics and analysis system, allowing the technical team to quickly and clearly understand the weak scenarios of the current intelligent driving system, and then optimize the intelligent driving algorithms more targetedly. The data loop enables Yuanrong Qixing to perform hundreds of technical iterations in a short time, creating high-level intelligent driving solutions without driving area restrictions.
Currently, all players are beginning to recognize the importance of data. Autonomous driving players will collect data related to autonomous driving, while automakers will collect driving-related data, and even some suppliers will establish their own data systems.
▲ Large-scale lane-level simulation
Although the automotive industry generates a lot of data, the current practice is that each player keeps their data separately and does not open it to the outside.
This leads to a problem where the data collected by different automakers may be quite similar, and the data collected by different autonomous driving players may also show basic similarities, resulting in everyone doing the same thing.
This results in some data being redundantly collected, while some valuable data cannot be utilized effectively, leading to the formation of data silos over time.
Wu Xuesong, Deputy General Manager of Changan’s Forward-looking Technology Research Institute and General Manager of Changan Zhitu, believes that to achieve rapid development of smart cars, it is essential to break down data silos and require more collaboration among companies for true development.
Companies like Tencent are also committed to breaking down information barriers.
In April of this year, Tencent shared its thoughts on “Car-Cloud Integration,” by transforming the R&D and production systems in the cloud, reconstructing the “software development and deployment” production line to improve R&D engineering efficiency; at the same time, building a car-cloud integrated data-driven operation platform, helping automakers build a user-operated “living map” that fully leverages data value to continuously enhance the in-car experience and expand innovative value-added services; and then extending connections to broader life scenarios through interactions between mobile phones and cars, and between cars and cars, continuously innovating user service models.
It is not difficult to see that Tencent’s “Car-Cloud Integration” concept aims to better and more efficiently leverage automotive data value, thereby creating more competitive products, aligning with the thoughts of industry experts.
Overall, the future of the smart car industry requires more players to participate together.
At this juncture, the automotive industry is undergoing the biggest upheaval in a century, and autonomous driving is undoubtedly a core area within this.
After more than a decade of exploration, autonomous driving and artificial intelligence technologies have developed a very close cooperation, and policy support is gradually becoming apparent. Tencent’s proposed digital twin + AIGC solution precisely aligns with the technological requirements for industry development.
As this solution continues to be promoted, true autonomous driving will accelerate its implementation.
