Xiaomi Automobile has recently welcomed a significant enhancement in its technological field——Former TuSimple China CTO Wang Naiyan has decided to join the team. Wang Naiyan will report directly to Ye Hangjun, the chairman of Xiaomi’s Technical Committee and the initial person in charge of Xiaomi Automobile's autonomous driving project. Before joining Xiaomi, Wang Naiyan held R&D and management roles at TuSimple China and was committed to developing L2-level assisted driving and L4-level autonomous driving solutions. He obtained his PhD from the Hong Kong University of Science and Technology, specializing in deep learning, and has been active in the field as a core developer of the open-source framework MXNet, publishing over 40 papers in top academic conferences and journals.

The news of Wang Naiyan's joining is seen as an important signal for the accelerated development of Xiaomi's intelligent driving technology. Xiaomi's founder, Lei Jun, announced that just 43 days after starting deliveries, the delivery volume of Xiaomi SU7 had exceeded 10,000 units, with the annual production capacity rising to 100,000 units. Lei Jun further indicated that the investment in intelligent driving technology would continue to increase. It is reported that Xiaomi's investment in the field of intelligent driving has exceeded 4.7 billion yuan, with a team of over 1,000 people, and their test mileage has exceeded 10 million kilometers. Xiaomi's urban NOA function is expected to be rolled out to users in 10 cities starting from May.

While Xiaomi is strengthening its investment in the field of intelligent driving, the group is also continuously seeking industry talents. In 2021, Xiaomi acquired the autonomous driving technology company DeepMotion for a price of 77.37 million US dollars, bringing its more than 20 core R&D members under its wing, which accelerated the pace of intelligent driving technology development. Wang Naiyan’s joining is considered to be helpful for Xiaomi to make greater progress in cutting-edge technologies in intelligent driving. Industry insiders evaluate that Wang Naiyan has a deep understanding of perception, planning control, and the entire chain of intelligent driving, and has his unique insights and understanding of the currently popular end-to-end technology solutions.

End-to-end technology is an important concept in deep learning, which allows AI models to directly produce final results from raw data. Tesla is at the forefront of applying end-to-end technology in intelligent driving, but Wang Naiyan has publicly stated that one should avoid falling into fixed thinking about Tesla's narrow interpretation of end-to-end and pointed out that end-to-end technology is one of the promising technical paths to solve advanced autonomous driving problems. However, he also cautioned that implementing this technology is still facing many challenges and is not an uncontroversial correct path.

In the field of intelligent driving technology, the focus of end-to-end is on achieving lossless information transmission, and fixed input and output points of signals may add unnecessary complexity and burden to the system. Wang Naiyan's unique insights and joining Xiaomi's intelligent driving team have injected new momentum for the company to overtake others. At Xiaomi's investor conference in 2024, Lei Jun stated that the company would further increase its investments in the intelligent driving field. The plan is to expand the intelligent driving team to 1,500 people this year and further increase it to 2,000 people by 2025, while investing about 1.5 billion yuan annually in intelligent driving technology.

Nowadays, the talent landscape in the intelligent driving industry is undergoing tremendous changes. XPeng Technology announced at a recent Tech Day event that its intelligent driving team will add 4,000 engineers, while on the other hand, Li Auto has started laying off its intelligent driving team. The addition of Wang Naiyan signifies the beginning of Xiaomi's intelligent driving strategy, and Lei Jun, with his strong determination and appeal, is expected to rapidly make up for the shortcomings in intelligent driving technology.

In the crucial battle of intelligent driving, artificial intelligence (AI) and data have become key elements of transformation. In traditional methods, stages like perception, decision-making, and planning control all require engineers to manually write code to set rules, which limits the scope and reliability of available scenarios. In 2021, Tesla introduced BEV (Bird's Eye View) technology based on the Transformer to the perception stage, achieving a shift from 2D images to 3D scenes. This improvement meant that intelligent driving no longer relies on high-precision maps, leading China's intelligent driving technology into the "light map era".

Further technological innovations, such as the launch of the Occupancy network, not only address the problem of pure visual perception's insufficient recognition of the depth information of objects on the road, but also improve the neural network's process from "knowing" to "recognizing," thus greatly expanding the application scope of pure visual solutions. These technological advancements have resulted in Transformer+BEV+Occupancy becoming the mainstream vision perception solution for the industry. Whether domestic manufacturers can quickly implement this technology solution has now become the key to competing for dominance in intelligent driving technology.

After Tesla completes the evolution of its perception module in 2024, it will further introduce AI neural networks in the decision-making and planning control fields, leading to the breakthrough of a so-called "end-to-end" large model. Tesla stated in its updated user manual that the FSD (Full Self-Driving software) has upgraded the technical stack for urban road driving to a single end-to-end neural network, trained on millions of video clips, replacing over 300,000 lines of C++ code. With the practical use of "end-to-end" neural networks, FSD has moved from Beta testing to a Supervised version and can provide services to millions of North American users.

XPeng Motors closely followed Tesla's technical roadmap and in May 2024, launched its Tianji System XOS 5.1. This system consists of a series of neural network technologies, including the XNet neural network, XPlanner planning control large model, XBrain large language model, etc., forming a complete end-to-end large model. XPeng expects to achieve nationwide road access in the third quarter of 2024 and to make a significant leap in urban intelligent driving experience in 2025.

Huawei continues to implement the use of LiDAR technology in its automatic driving system ADS 3.0 and has made more technical optimizations on this foundation. The latest system has abandoned the previous use of the BEV network and switched to using the GOD network for perception processing, utilizing the PDP network to execute pre-decision and planning. The application of these technologies, especially with the addition of LiDAR, has further enhanced Huawei's performance in the Active Emergency Braking (AEB) aspect, with notable effectiveness.

Meanwhile, Xiaomi revealed at its product launch event the application of end-to-end neural network technology in parking, achieving significant results. The technology can now precisely park in extremely narrow spaces with an accuracy of up to 5 centimeters and can perform valet parking at a speed of 23 kilometers per hour. Although the development of end-to-end neural network models is just beginning, the training of massive amounts of data is a crucial factor in the key role of this technology. Huawei's currently hot-selling Wenjie models have provided it with a wealth of data support, in contrast to the powerful tech-driven XPeng Motors which recently faced challenges of declining sales.

Despite starting a bit later in the development of end-to-end technologies, Xiaomi, with the breakthrough of 88,000 units sold within 33 days of the market launch of its first model and an impressive 82.39% activation rate of intelligent driving, is quickly emerging as a significant competitor in the data field not to be overlooked.