The head of a smart driver will grow into the ideal shape

Li Auto did not put too much energy and R&D capabilities into expanding the L6 product line, allowing Li Auto to free up time to engage in other activities. At this time, Li Auto needs to make some changes. At the Shanghai Auto Show, we witnessed Ideal demonstrating its 800V high-voltage pure electric platform and successfully building a 480kW/240kW high-rate self-operated fast charging equipment. Such a seemingly simple but very practical project is very important and must be done for Li Auto. For those who pursue their ideals, this is a matter of vital importance, not an emergency.

如此紧急且至关重要的事情，我们必须予以高度重视？这意味着，它将在无人驾驶领域取得突破。当然，李想本人在年初的公开信中表达了对于汽车成为2030年领先的人工智能企业的美好愿景。这个目标已经达成了。在2023年，这个“领先”成功实现了城市场景的导航辅助驾驶，为城市交通带来了显著的效益。与高速场景不同，几位头部甚至非头部玩家在同一时间段内同时着陆，竞争势必更加激烈。

在今年的推荐中，理想的选择是将他们的城市NOA功能推送至年底，覆盖100个城市，以达到更广泛的覆盖面。理想汽车是怎样做到这一点的？这位家用车专家怀揣着成为人工智能领域的先锋的决心，他们与友商所推出的产品在某些方面有着惊人的相似之处，但在其他方面则存在着显著的差异？在智能互联时代，理想汽车如何突破自我？在技术路线的不断演进中，理想汽车所具备的优越性何在？

We explored these issues in our conversation with the two heads of Lili Intelligent Driving—Vice President Dr. Lang Xianpeng and Product Leader “Squad Leader” Zhao Zhelun. I think these two people have very strong ideas and a lot of practical problems that need to be solved. Although these problems are not complicated, in the process from the bottom layer to practical application, we can gradually understand their deep nature.

BEV is a network model that uses graph-free iterative GPT to implement...

Like most of its peers, Li Auto uses a series of well-known tools to perceive and evaluate the underlying principles of the world.

The terms "BEV" and "Occupancy" were first encountered by our counterparts. Here I would like to briefly introduce the relationship between them, hoping to bring some inspiration to readers. The former refers to the "bird's-eye view", which uses various vehicle sensors to accurately perceive the surrounding environment and generates a "God's perspective" environmental simulation; the latter refers to the "intelligent driving system", which can use environmental information based on React so that drivers get to their destination faster and better. Occupancy, also known as "network occupancy", is a method of further refining environmental simulation to help vehicles identify the difference between obstacles and passable road surfaces, thereby providing the second step of "perception-decision-execution" provide data. For driverless driving, these two words are very directional, because they both involve the relationship between the car and the road, and how people drive the car. Among the three major elements of artificial intelligence, algorithms, computing power and data form the ideal components.

The "static BEV" elements in BEV, including road structures and markings, form part of high-precision maps in the traditional sense; such as parking lots and other facilities and surrounding road network information, etc., and their functions are mainly traffic guidance and driver assistance. Provide reference basis. Correspondingly, "dynamic BEV" refers to moving vehicles, pedestrians, etc., which together form a dynamic transportation network. With the advent of the smart car era, the traditional "static" or "dynamic" concepts will be gradually eliminated. Considering the cost and freshness of high-precision map collection and the development of in-vehicle algorithms, we recommend that, like other leading manufacturers, let vehicles independently build static and dynamic BEV models, thereby reducing the need for navigation in smart driving to navigation maps + A little level of intelligent driving elements, i.e. only need to know the destination and path, without knowing "trying to merge left at XX latitude and longitude".

For those non-standard objects that may affect driving, the Occupancy network provides an ideal method to restore the entire world to a Minecraft area full of blocks, where the vehicle can identify which obstacles require braking and which roads can be passed. , and where the "magic carpet" suspension needs to be softened.

The operation of these models cannot be separated from the support of computing power, which is an indispensable prerequisite. If a suitable method can be found to solve this problem, then the future of self-driving cars will definitely become faster, farther and safer. Therefore, we may then face a situation where the presence of an object is observable on the visualization interface, but the system is unable to respond to it. The AD Max version is equipped with two NVIDIA Orin chips, which makes the ideal state of urban NOA possible because it can differentiate based on computing power. However, next we will try to migrate some of Max's computing models to the Pro version to enable their sharing on the intelligent driving technology stack. For example, this year, they will use the BEV perception architecture to rewrite the high-speed NOA and LCC functions.

Do you remember the three basic elements of artificial intelligence? Where does the data come from? According to Dr. Lang Xianpeng, although there are minor differences in algorithms and computing power, data is the key to widening the gap. What happens to data after it is machine learned? What is the method to get the data? These all require a well-supported platform. What will be the subsequent operations for the obtained training iterations? This is an ideal area of ​​strength.

The history of collecting intelligent driving data can be traced back to   the  era of Ideal One  ,   and the first algorithm training video was shot in 2019. At that time, Ideal was still developing autonomous vehicles. With the advent of the AD1.0 era, the 2021 Ideal One has begun production, and they have collected representative data of up to 100 million kilometers. According to the data formula of "intelligent driving hardware Through continuous iteration, Ideal finally formed a complete road model library. So far, our ideal algorithm training mileage has reached an astonishing number of 400 million kilometers.

The collected data is fed back to support algorithm development. This is "accumulation of experience". In addition to conducting actual road tests, we have also stocked up a large number of high-power chips and established our own large-scale supercomputing center to quickly fight monsters and accumulate experience in the simulation environment. This is like how we used mechanical tools to do some simple and repetitive things, and then fed the results back to the system for correction and optimization. Similar to the experience accumulation method of today's ChatGPT, the basic capabilities of urban NOA are obtained through a large number of weakly related and roughly labeled samples, and then improved through individual precisely labeled data (including long-tail problems and unexpected problems in individual intelligent driving) . It’s clear that today’s vehicles are able to think about roads and trajectories in a human way, rather than what we previously thought were mechanized sea-going tactics.

As users, what valuable information will we discover?

The issues involved in the application layer are more complex and intuitive than other layers. We will personally experience the practicality of an urban navigation assisted driving system, including the frequency of human takeover, scope of application and most importantly, cost issues. We've got the latest news from two of the world's leading smart driving experts, whose expertise and experience has benefited us greatly.

Divide manual takeovers into two categories, one is to provide experience, and the other is to ensure safety. In order to achieve this goal, Ideal formulated a specific plan in the early stages of development - to reduce the number of takeovers by increasing system pressure. Considering the user's awareness of the boundaries of system capabilities, there are no special requirements for "experience takeover". However, for "safe takeover", its use is required to be reduced as much as possible. Comprehensive consideration, based on the test data of Li Auto, the current version of Urban NOA can basically take over once every 20 to 30km. The next step is to divide the time into once a day and once a week. However, in reality, due to factors such as road congestion, "experience takeover" will also cause frequent switching. We can use the urban NOA function with confidence because this frequency has achieved the desired effect. To ease the discomfort caused by "experience takeover," we plan to enhance the vehicle's visual interface so that drivers can clearly understand the system's capabilities and future plans.

As far as the scope of urban NOA is concerned, the most ideal car is also determined based on the number of vehicles and the number of trainings. For those road sections and cities with a large number of cars and frequent driving by car owners, the opportunity to open urban NOA earlier has arrived. For drivers with certain driving experience, "HD Map" is a good choice. The ideal of every vehicle equipped with AD Max is a "high-precision mapping vehicle" and an "intelligent driving algorithm training vehicle."

In the rapid rise of smart driving, data plays a vital role, so it is ideal to regard smart driving as a product rather than a service, so that when users sell their vehicles, data and algorithms become an important part of the vehicle's assets and residual value. part. Ideal believes that smart travel requires a complete closed loop, including vehicle usage data, vehicle trajectory data, and driver personal preference data, which includes predictions of vehicle performance. For this reason, the most ideal option right now is to offer AD Max users free shipping in city NOA. This is provided by a project called Smart Drive, which aims to help car manufacturers and operators optimize their operating strategies to improve profits by using data. They are eager to apply AD Max in the closed loop of smart driving data to improve the value preservation ability of user vehicles, and are more inclined to use it as a component of smart driving data.