Catch up with the first echelon in three years, Xiaomi Intelligent Driving’s latecomer advantage

"High-speed NOA will be available for the entire series when it is launched, and urban NOA will be available nationwide in August this year." At the launch conference of Xiaomi SU7, compared with the interior and smart cabin, founder Lei Jun talked about autonomous driving in a shorter space, but the content was very important - in three years, starting from scratch, how can Xiaomi achieve intelligent driving capabilities comparable to those of the first echelon? We have combined the information recently disclosed by Lei Jun and others, combined with past public information, and disassembled from multiple dimensions such as personnel, software and hardware , and development process , trying to review the process of Xiaomi's intelligent driving from 0 to 1.

01. Latecomer advantage: catching up with the first echelon in three years

Three years later, the Xiaomi Intelligent Driving team has delivered a satisfactory result. Like most car companies, Xiaomi Intelligent Driving has two solutions: Xiaomi Pilot Pro and Xiaomi Pilot Max.

Pro is a pure vision solution, and Max is a vision + lidar solution. Although the technical configurations are different, they are both equipped with 11V high-definition cameras , NVIDIA DRIVE Orin chips , and full-stack self-developed algorithms.

Among them, the PRO version is equipped with the Xiaomi SU7 standard version and can support high-speed road navigation assistance (NOA). The MAX version can also realize urban navigation assistance (NOA) based on the functions of the PRO version. The PRO version is equipped with 84TOPS computing power . According to Lei Jun's description, this is enough for use. AEB can stop at a speed of up to 135km/h, which exceeds Tesla . If the national city NOA can be launched in August this year as expected, then Xiaomi Auto will be well-deserved in the first echelon of intelligent driving.

After all, only Huawei has achieved "all-around driving" at present. Xiaopeng Motors' urban XNGP is currently only available in 243 cities across the country, and "all-around driving" is limited to some users.

At the first launch of Xiaomi Auto, Lei Jun demonstrated three autonomous driving scenarios: parking, urban area and highway. The most impressive demonstration was that the Xiaomi SU7 drove automatically to the 7-story open-air parking lot in Dongguan without anyone in the parking lot, including a series of actions such as turning back, avoiding pedestrians, finding a position, and landing.

According to official statements, Xiaomi is the first car manufacturer in China to apply the end-to-end big model to the automatic parking function, integrating perception, decision-making, and planning algorithms into one. It is also the most human-like intelligent driving algorithm to date, which can achieve 5cm accuracy in extremely narrow parking spaces and valet parking at a cruising speed of 23km/h.

In terms of safety, Xiaomi claims that even the lowest-configuration model, Xiaomi SU7, can be fully taken over less than once in 300 kilometers, and there will not be a single sudden acceleration or brake in 100 kilometers.

In terms of practical experience, Lei Jun said that he drove 1,270 kilometers from Beijing to Shanghai, and only took over three times in the middle. From Lei Jun's point of view, whether it is the speed of implementation or the practical level, Xiaomi's intelligent driving is at the top of the industry, and it is supported by a full stack of self-developed algorithms, namely: self-developed underlying algorithms, large road models and occupied networks.

Like most first-tier automakers, Xiaomi's intelligent driving system algorithm uses BEV+Transformer+OCC, which is also the most efficient solution at present. Among them, the Xiaomi intelligent driving team has made innovations in both BEV and OCC dimensions.

In terms of BEV, Xiaomi cars are equipped with zoom technology. According to the explanation in the press conference PPT, the more accurate the understanding, the higher the resolution. The output resolution of BEV is determined by the parameters at the time of input. If BEV can achieve zoom, it means that the parameters at the beginning are fluid.

In other words, Xiaomi Intelligent Driving has prepared multiple sets of output models that can be switched according to the scene. In the occupancy network part, the Xiaomi team jointly published a paper with Peking University in 2023, "Unified Vision-based 3D Occupancy Prediction with Geometric and Semantic Rendering", which provides a technical path for "Uni-OCC". The highlight of Uni-OCC is that it reduces the dependence on 3D pixel recognition and greatly reduces the recognition cost.

At the press conference, Lei Jun compared Xiaomi's OCC technology to the basic blocks in the game "Minecraft". If the traditional occupancy network is a "big block", then Xiaomi's OCC is a "small block". The accuracy of the super-resolution occupancy network model has reached a level of less than 0.1m, while the occupancy network accuracy of Tesla FSD is 0.32m.

In addition to improving the underlying algorithms, Xiaomi also has some new ways to play. For example, the self-developed road model technology. Officials say that the real-time generated road topology effect is comparable to high-definition maps. In actual scenarios, when the road ahead is under construction, the road model can also accurately generate intelligent guide lines, predict in advance, and bypass smoothly.

In addition, Xiaomi has also developed a technology for automatic noise reduction in rainy and snowy days. The original input data is processed by an algorithm before entering the network, and the outline of objects can be accurately identified even in unconventional weather. Of course, just as every car company wants to be on par with Tesla in intelligent driving, Xiaomi has also considered adopting a more radical pure visual intelligent driving solution.

According to Xiaomi's core technical personnel, Xiaomi Auto considered a pure vision solution in 2021, but later adopted a lidar + visual assistance solution. If Xiaomi is in the leading position in terms of software , then its deployment at the hardware level is at the mainstream level.

In terms of chip configuration, there is a gradient of chips among the three models of Xiaomi SU7. Xiaomi has installed two NVIDIA Orin X chips in the high-end models, which is at the same level as the mainstream domestic new energy models.

Some smart electric vehicles have the most mature and advanced intelligent driving chips and the latest technology stack. Thanks to this, Xiaomi, which entered the market late but has the "latecomer advantage", has also avoided many detours. At the same time, except for Huawei, a "car company that does not build cars", Xiaomi Auto is the only car manufacturer that has not looked for any intelligent driving suppliers in the field of autonomous driving. This also makes it more proactive and able to keep up with the iteration of intelligent driving technology routes quickly.

02. Started from AI lab and made long-term commitment

Xiaomi Motors has decided to develop its own full stack since its establishment. The initial team came from Xiaomi Group AI Lab, and the team is expanding rapidly. On September 1, 2021, Lei Jun released a photo of the Xiaomi Motors startup team.

The sixth person from the right in the second row is Ye Hangjun, who created the "Little Ai classmate" photo. The first row from left to right are Lin Shiwei, Zhang Feng, Liu De, Wang Xiang, Lei Jun, Hong Feng, Lu Weibing, Qi Yan, and He Yong; the second row from right are Qin Zhifan, Liu Anyu, Li Tianyuan, Yu Cuo, Li Xiaoshuang, Ye Hangjun, Chen Jinhong, and Fan Jialin.

Most of these people are transferred from Xiaomi, and most of them are from the marketing and government relations departments. The only one who is related to the field of intelligence is Ye Hangjun, who created "Xiao Ai Tongxue". Ye Hangjun graduated with a doctorate from the Department of Computer Science at Tsinghua University , and his research field is computer vision and image retrieval.

Before joining Xiaomi, he had worked in the field of Internet search for a long time. From 2006 to 2010, Ye Hangjun worked at Google. Two years later, Ye Hangjun joined Tencent and worked on the technical research and development of search engines.

In 2012, Ye Hangjun chose to switch to Xiaomi Group, which is more manufacturing-oriented. Facts have proved that Ye Hangjun's career is in line with the development of the times. After 2012, search engines were eliminated by information flow.

Since 2016, AI has become Xiaomi's core strategy for the next decade. At that time, Ye Hangjun was the general manager of the Artificial Intelligence Department. When the Artificial Intelligence Department was established, Ye Hangjun "recommended himself". Later, he was also appointed as the vice chairman of the AIoT Strategic Committee. Five years later, he was promoted from general manager of the Artificial Intelligence Department to chairman of the Xiaomi Technology Committee. In early 2021, he was "appointed" by Lei Jun to preside over Xiaomi's autonomous driving business.

Ye Hangjun once said that at the beginning, the entire Xiaomi team did not have a deep understanding of the field of autonomous driving. Although there was an economic basis to "let go and do it", the most important thing was to have a strong decision-making and execution team, which required a lot of manpower investment.

One year after Xiaomi announced its car manufacturing plan, in mid-2022, Ye Hangjun said in an interview with LatePost that the intelligent driving team had more than 600 people at the time, of which more than 70% had master's and doctoral degrees, covering all directions of intelligent driving software and hardware. The core of autonomous driving is artificial intelligence, so most of the intelligent driving team came from internal AI laboratories. In just one and a half years, Xiaomi's intelligent driving team has reached such a scale, which shows how much investment it has made.