Frame-by-frame analysis of Xiaomi's self-driving capabilities

I'm coming! Xiaomi car, the most critical core progress: autonomous driving capability, publicly demonstrated for the first time!

Whether you are bragging or self-confident, Lei Jun has already made a reservation: autonomous driving will rush to the first echelon in 2024. In the end what kind of ability? The key highlights are analyzed in detail.

8 minutes, what did Xiaomi show?

There are three areas of high speed, city, and parking, and there are many bright spots.

The first is the high-speed scene. Basic L2 functions such as lane keeping and cruise control are unremarkable. The highlight of Xiaomi's autonomous driving is that it has broken through the simple "assistance", and already has the functions of actively selecting lanes, entering and exiting ramps, and merging and overtaking.

Self-directed up and down high-speed - behind the high-precision map

In high-speed scenarios, vehicles can get on and off ramps autonomously.

This single function actually represents that Xiaomi’s autonomous driving adopts the solution of connecting with high-precision maps. Micro-operations during vehicle driving, such as avoidance, braking, direction, etc., are completed by onboard sensors and computing platforms.

However, the decision-making capabilities of route planning, entering and exiting intersections, and going up and down high speed are inseparable from the support of high-precision maps.

Of course, high-precision map information is to provide the system with large-scale route planning and road emergencies. In actual driving, sensors and positioning modules are still needed to "tell" the system what kind of scene it is in.

This also shows that Xiaomi has also made progress in the fusion of different sensor data.

High-speed autonomous merging and overtaking

This part can actually be split into two actions. The big level is that the system AI will autonomously select the fast lane to pass in high-speed scenarios. That is to say, after entering the expressway from the ramp, it must be accompanied by the action of changing lanes and overtaking.

However, the fast lane strategy has not been disclosed, that is, it is not clear how long in advance the Xiaomi automatic driving system will prepare to change lanes and leave the high speed, and it is also unclear how the system will react when there are low-speed targets in the fast lane.

In the process of changing lanes, it is inevitable to encounter unexpected situations, and the avoidance strategy is very critical:

It can be seen from the displayed scene that when the system is about to change lanes, there are low-speed large vehicles in front and fast continuous lane-changing vehicles chasing after them.

From this case alone, the strategy of Xiaomi's automatic driving system is very appropriate: first, slow down slightly to open the distance between the cars in front, and then wait for the rear car to pass before changing lanes.

The whole process did not exit the takeover, nor did it slam on the brakes, slam the direction, and so on.

This also shows that the system is more "old driver", not blindly cautious, nor too aggressive a driving style.

The hardest unprotected turn: U-turn

The highway conditions are relatively clear and single, and the test is more of a strategy. The real test of perception, recognition, and prediction is the urban road.

In Xiaomi's demo, there is a scene that has not been demonstrated by a friend before - an unprotected and autonomous U-turn:

The difficulty here is that the target lane is not blocked by traffic lights, and there are vehicles passing through at all times. Therefore, at the data level, not only the forward-facing camera is required, but also the ability to look around.

In the background of the system, it is necessary to restore and reconstruct the 3D scene from the 2D data of several different cameras.

Avoid target

There are two types: avoiding stationary targets and passing through complex intersections.

The first is to bypass the accident car occupying the road:

The system can also pass through complex intersections where people and vehicles are mixed and there are no traffic lights:

However, the braking action here is obviously more violent than the high-speed scene.

At intersections with traffic lights, the system can also take reasonable actions based on traffic signals:

Automatic entry and exit roundabout

This is a roundabout with a larger radius and a gentler curvature.

Getting in and out smoothly is actually not a very difficult challenge now.

What really tests the technology is the system's construction of the entire environment. The core is whether the sensor accuracy can detect various central roundabout roadbeds with lower heights and different colors, or the system can control the speed and attitude of vehicles in curves with different curvatures. Ride comfort cannot be guaranteed.

And these, I am afraid that we will only know until the day of mass production and delivery.

However, the roundabout scene also has a unique meaning. Entering and exiting the roundabout, including the U-turn behavior in front and the recognition of traffic lights, all indicate that the Xiaomi automatic driving system also realizes high-level navigation assistance functions based on high-precision maps in urban scenarios.

It has the same nature as Xiaopeng City NGP, Momo NOH, Tesla NOA, etc., which will be launched in the second half of the year.

Self-parking and self-charging

What Xiaomi shows is not the conventional identification of parking spaces, but the memory parking function that starts from the entrance of the parking lot:

That is, the system needs to perceive the entire parking structure first, then build a digital twin and remember the route to the parking space.

In the process of traveling, it is necessary to make real-time avoidance of passing vehicles and people.

From the HMI interface, it can be clearly seen that the system has accurately identified different targets such as cones, vehicles, columns, motorcycles...etc.

Moreover, on the main target, the system also indicates the distance and speed information from the vehicle in real-time.

It is not too difficult to identify stationary targets at low speeds, but it is still necessary to fuse data from different sensors such as cameras, lidars, and millimeter-wave radars.

And due to the light conditions of the basement, the redundant role of the radar will be more prominent.

There is nothing unusual about reversing the car into the garage, but the difference is that Xiaomi showed a car "charging treasure" that can move freely:

What's wrong?

Xiaomi autonomous driving cannot be perfect. Through the previous analysis, we already know that many scenarios are not passed even if they are qualified, and Xiaomi has not disclosed the difficulties of the real test of the technology.

For example, in the scene of avoiding a faulty car, there is actually doubt. The detour shown in the video is done with the dotted line. In fact, this road is only a dashed line in the section of the borrowed road, and the "faulty car" happened to break down here, and it did not occupy the entire lane. In the case of solid lines or even double yellow lines, how to deal with Xiaomi is not shown.

In addition, the shortcoming that can be seen at a glance is that the "cautious" of Xiaomi's automatic driving system is far greater than "confidence" on urban roads, and the following distance is too long to plug a light truck. This also caused Xiaomi's self-driving test cars to be frequently jammed on the road.

Moreover, in the event of a sudden stoppage, pedestrian crossing the road, etc., there will be a more violent brake, and then the vehicle will be delayed for a second or two before continuing to drive.

But in any case, the autonomous driving shown by Xiaomi does not have a low starting point. High-speed pilot assistance, city pilot assistance, and autonomous parking all have outstanding performance. Among them, high-speed pilot assistance, first-tier OEMs and autonomous driving companies have already mass-produced and installed, and the focus of competition in the second half of this year is urban pilot assistance.

As for autonomous parking, each company has different levels of experience. Xiaomi's ability to display is at least on par with Xiaopeng Motors. The integration of these three domains is the filing index set by most companies in the industry, and the time point is 2025.

If the mass production of Xiaomi Auto in 2024 will follow step by step, it means that Lei Jun has already planned the debut of Xiaomi Auto, namely L3, to compete with new forces, Huawei and Tesla on the high-end battlefield.

Xiaomi autonomous driving, what plan? Where did it come from?

Xiaomi's self-driving test car uses the BYD Han EV in the open road scene. The parking scene uses a BMW 3 Series. It is only used as a technical verification to provide a wire-controlled chassis for Xiaomi.

For the specific plan, what can be seen is that the body is covered with cameras all around, covering 360°, and there is a raised lidar on the roof. The rest of the sensors and underlying computing power are unknown. Lei Jun said that this is the result of Xiaomi Auto's 500 days, but this is actually a bad statement.

The core team of Xiaomi Auto's autonomous driving comes from DeepMotion, an autonomous driving startup acquired last year.

DeepMotion was founded in 2017, and the core team came out of Microsoft Asia Research Institute. When it became a subsidiary of Xiaomi last year, it is said that it has successfully developed technical solutions such as multi-sensor fusion perception modules, the automatic composition of high-precision maps, and high-precision positioning based on vision.

Therefore, Xiaomi's autonomous driving is actually in line with the objective laws of the industry, and it is also the accumulation of achievements of four or five years of research and development.

After the debut, what kind of reaction did it get?

There are many doubts, and they are merciless under the official Xiaomi account:

The main doubt is that most of the autonomous driving of the new forces now comes from the open-source of Baidu Apollo.

In addition, Xiaomi's self-driving HMI is also questioned to be rendered in a later animation:

Of course, there are also appreciations. But most of all, I still expect Lei always to bring young people their first smart car.

Xiaomi car, Lei Jun's latest spoiler

The actual progress of Xiaomi Auto, including talents, funds, patents, etc., has been introduced in detail before the smart car reference. This time we only look at the new information from Lei Jun himself.

For the first time, Lei Jun officially disclosed the preparation details of some technologies and personnel.

Specifically. Xiaomi invested 3.3 billion yuan in research and development in the field of autonomous driving in the first phase and established a test fleet of more than 140 vehicles.

At present, the team has more than 500 people. At the same time, through wholly-owned mergers and acquisitions and industrial investment in upstream and downstream enterprises, it promotes the medium and long-term industrial layout in the field of autonomous driving. We may be concerned about "the first car for young people", but Mr. Lei doesn't see it that way.

"Crossing the trough and laying the foundation for reputation", Lei Jun believes that autonomous driving is the core competitiveness.

Lei Jun revealed that since Xiaomi announced its entry into the smart electric vehicle industry, it first entered the field of autonomous driving, which is the most technically difficult. Get started.

By the end of this year, the overall size of the autonomous driving team is expected to exceed 600 people, of which 50 top experts form the backbone team, with Boshuo accounting for more than 70%, covering sensors, chips, perception regulation algorithms, simulation technology, high-precision maps, high-precision Talents required for full-stack autonomous driving technologies such as accurate positioning, toolchain, and training capabilities.

Most of the core backbone members have backgrounds such as Microsoft and Google.

In Lei Jun's annual speech, he put the great cause of car manufacturing and autonomous driving into "One more thing", which is obviously the highlight of Xiaomi's future.

According to the information obtained by the smart car reference, Lei Jun's current energy is almost all invested in the Xiaomi car project. The core mobile phone business is now only providing guidance at the strategic and planning level, and Zeng Xuezhong is fully responsible for the specific work.

For car building, Lei Jun does have a tendency to "break the boat and sink the boat". After the progress of autonomous driving, Lei Jun said that in the next two years, Xiaomi Auto will no longer disclose the progress, and will no longer refute rumors and rumors. All suspense, 2024 delivery announced.