Jia Haonan Posted from the copilot temple

Reference for AI Auto | Official Account AI4Auto

Here it comes!

Xiaomi’s most crucial core progress in the automobile industry: Autonomous Driving capability is publicly displayed for the first time!

30 highlights of capabilities, Lei Jun summarized it as Xiaomi’s initial answer sheet for 500 days:

Whether it’s boasting or confident, Lei Jun has already set a goal: to rank first in autonomous driving by 2024.

What kind of ability is it? Here is a detailed analysis of the key highlights.

8 minutes, what did Xiaomi show?

There are three domains: highway, city, and parking, and there are many highlights.

First of all, in the highway scene, basic L2 functions such as lane keeping and adaptive cruise control are ordinary.

Xiaomi’s autonomous driving highlights lie in the breakthrough of simple “assistance” and it has realized functions such as active selection of lanes, entering and exiting ramps, and lane-changing overtaking.

Autonomous access to and exit from the highway – backed by high-precision maps

Under the highway scene, the vehicle can achieve autonomous access to and exit from the ramp.

Behind this single function, it actually represents Xiaomi’s autonomous driving solution, which is connected to high-precision maps. The micro-operations during the driving process, such as avoiding, braking, and direction control, are all completed by the sensor and computing platform on the vehicle.

However, the decision-making level of route planning, entering and exiting the intersection, and access to and exit from the highway requires the support of high-precision maps.

Of course, high-precision map information is used to provide large-scale route planning and sudden changes in the road. In actual driving, sensors and positioning modules still need to “inform” the system of the scene it is currently in.

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

Autonomous lane changing and overtaking on the highway

This part can actually be split into two actions.

On a larger scale, the system’s AI will autonomously choose the fast lane for passage in a highway scene. That is, when entering the highway from the ramp, there must be a lane changing and overtaking action.

However, the fast lane strategy has not been revealed. It is still unclear how long Xiaomi’s autonomous driving system will prepare to exit the highway in advance or what kind of reaction it will make when the fast lane has a low-speed target obstacle.In the process of changing lanes, unexpected situations are inevitable and avoidance strategy is key:

From the displayed scene, it can be seen that the Xiaomi autonomous driving system’s strategy is very appropriate: first, slightly decelerate to increase the distance to the front car, and then wait for the rear car to pass before changing lanes.

The whole process does not involve taking over control, nor does it involve emergency braking or sudden steering.

This also shows that the system is quite experienced and does not avoid risks rashly, nor has an overly aggressive driving style.

The Most Challenging Scenario: U-turn without protection

The high-speed road condition is relatively clear and simple, testing the strategy more.

However, the real challenge for perception, recognition, and prediction lies in the urban roads.

In Xiaomi’s demo, there was one scenario that had not been demonstrated by other competitors —— a U-turn without protection:

The difficulty of this scenario lies in the fact that the target lane does not have traffic lights to stop the traffic, and vehicles pass by frequently. Therefore, not only does it require a front camera at the data level, but also panoramic vision ability.

In the system background, 3D scenes need to be reconstructed from 2D data of different cameras.

Avoiding Targets

Two types: avoiding stationary targets and the complexity of passing through intersections.

First, the accident vehicle blocking the lane:

In a complex intersection with mixed pedestrian and vehicle traffic and no traffic lights, the system can still pass smoothly:

However, the braking action here is clearly more severe than in the high-speed scenario.

At intersections with traffic lights, the system can also make reasonable actions according to traffic signals:

Automatic Entry and Exit of Roundabouts

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

Successfully entering and exiting is not a very difficult challenge now.The real challenge for technology lies in the system’s ability to construct the entire environment, with the core being whether the sensor’s accuracy can detect various central roundabout roadbeds that are even lower and have different colors, or whether the system can control the speed and posture of the vehicle in different curve roads to ensure comfortable riding. However, these may only be revealed on the day of mass delivery.

However, the roundabout scenario also has a unique significance. Entering and exiting the roundabout, including the U-turn behavior in front and the recognition of traffic lights, all demonstrate that Xiaomi’s autonomous driving system has also achieved high-level navigation assistance based on high-precision maps in urban scenarios.

Like the upcoming XPeng City NGP in the second half of the year, MoNOH, and Tesla NOA, etc., they all have the same nature.

Autonomous parking and charging

What Xiaomi demonstrated was not the usual recognition of parking spaces and parking, but a memory parking function that starts from the entrance of the parking lot.

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

During the process, real-time avoidance needs to be made for incoming and outgoing vehicles and pedestrian targets.

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

Moreover, the system also displays the distance and speed information between the primary target and the vehicle in real-time.

Identifying stationary targets at low speeds is not too difficult, but it still requires the fusion of data from different sensors such as cameras, LiDAR, and mmWave radar.

Furthermore, due to the lighting conditions in the underground parking lot, the redundant role of the radar will be more prominent.

There’s nothing special about reversing into the parking spot, but what’s different is that Xiaomi demonstrated a “power bank” car that can move freely:

What are the drawbacks?

Xiaomi’s autonomous driving system cannot be perfect.

Through the front analysis, we already know that many scenarios are not qualified even if they pass, and the real challenges of technology that Xiaomi has not revealed.

For example, in the scenario of avoiding a broken-down vehicle, there is a suspicion point.

As shown in the video, the bypass was completed under the dotted line. In fact, this road section is only a dotted line, and the “broken-down vehicle” happened to break down here, and it did not occupy the entire lane.

When facing solid or even double yellow lines, Xiaomi did not demonstrate how to handle it.The Markdown text translated into English Markdown text, with HTML tags preserved, and only the corrected and improved parts answered in concise and professional language:

Moreover, it is obvious that Xiaomi’s autonomous driving system is much more “cautious” than “confident” in city road conditions, with an overly long following distance, even allowing a light truck to cut into its lane.

This has caused the Xiaomi autonomous driving test car to be frequently cut in on while on the road.

Furthermore, when encountering sudden cut-ins or pedestrians crossing the road, there is a relatively sudden braking followed by a delay of one to two seconds before the car can resume driving.

But regardless, Xiaomi’s autonomous driving demonstration showcases immense potential.

The three domains of high-speed navigation assistance, city navigation assistance, and self-parking, all exhibit impressive performance.

Among them, high-speed navigation assistance has already been implemented in mass production by first-tier car manufacturers and autonomous driving companies, while the focus for competition in the latter half of this year is expected to be city navigation assistance.

As for self-parking, all companies have their own versions, varying in their experiences, but Xiaomi’s demonstrated capabilities are at least on par with those of XPeng Motors.

Integration of these three domains is the benchmark for most industry players, with the timeline set at 2025.

If Xiaomi Car can begin mass production as planned in 2024, it shows that Lei Jun has already planned for Xiaomi Car to compete at the L3 level in the high-end battlefield, alongside new forces such as Huawei and Tesla.

Xiaomi’s Autonomous Driving: What’s the Plan and Where Does it Come From?

The Xiaomi autonomous driving test car used on open roads is a BYD Han EV.

For parking scenarios, a BMW 3 Series is used, both only for technical verification to provide Xiaomi with line control chassis.

The specific plan includes cameras covering 360° around the car and a high-altitude Lidar on the roof, but the rest of the sensors and underlying computational power are unknown.

Lei Jun said that this was the achievement of 500 days of Xiaomi Car’s development, but this statement is incorrect.

Xiaomi’s autonomous driving core team comes from last year’s acquisition of the DeepMotion autonomous driving start-up.

DeepMotion was founded in 2017, and its core team came out of Microsoft’s Asia-Pacific R&D Institute. When it became a Xiaomi subsidiary last year, it reportedly had already developed technology such as multi-sensor fusion perception modules, high-precision map automation composition, and high-precision positioning based on vision.

Therefore, Xiaomi’s autonomous driving is in line with the objective laws of the industry and is the result of four to five years of accumulated development.

After the debut, what kind of reactions did it cause?

There were many doubters, and there was no mercy even from comments under Xiaomi’s official account.

The main criticism is that most new players in the autonomous driving field rely on the open source technology of Baidu Apollo.

In addition, Xiaomi’s autonomous driving HMI has also been criticized for being rendered with animation in the later stage.

However, there are also a lot of admirers. But most of them still hope that Mr. Lei can bring the first intelligent car to young people.

Xiaomi Automobile, the latest update from Mr. Lei Jun

The actual progress of Xiaomi Automobile, including talent, capital, patents, etc., have been detailedly introduced before.

This time, we only look at the new information disclosed by Mr. Lei Jun.

Mr. Lei Jun officially revealed some technical and personnel preparation details.

Specifically, Xiaomi invested 3.3 billion yuan in R&D expenses in the first phase of autonomous driving field, and set up over 140 test car fleets.

Currently, the team size exceeds 500 people, and it has promoted the medium-and long-term industrial layout in the autonomous driving field through wholly-owned mergers and acquisitions and industrial investments in upstream and downstream companies.

What we care about may be the “first car for young people”, but Mr. Lei Jun does not think so.

For the car-making business “crossing the low valley and staking reputation,” Mr. Lei Jun believes that autonomous driving is the core competitiveness.

Mr. Lei Jun revealed that since Xiaomi announced its entry into the smart electric car industry, it has firstly focused on the autonomous driving field with the highest technological difficulty, and realized the start from scratch through three major measures, including heavy investment, the construction of top teams, and the up-and-downstream layout of industrial investments.

By the end of this year, the overall scale of the autonomous driving team is expected to exceed 600 people, including a backbone team consisting of 50 top experts, with a proportion of more than 70% having doctor’s and master’s degrees, covering full-stack talents required for autonomous driving, such as sensors, chips, perception control algorithms, simulation technology, high-precision maps, high-precision positioning, toolchains, and training capabilities.

Most of the core members have a background in Microsoft, Google, and other well-known companies.

In the annual speech by Mr. Lei Jun, car-making business and autonomous driving were put in as “One more thing”, obviously treating Xiaomi’s automobile as the future heavyweight show.According to the Intelligent Car Reference, it is reported that Lei Jun is currently putting almost all of his energy into the Xiaomi car project. As for the core smartphone business, Lei Jun is only providing strategic and planning guidance, while all specific work is being fully conducted by Zeng Xuezong.

Regarding car manufacturing, Lei Jun is demonstrating a “break the boat and burn the bridge” determination.

Since the progress of autonomous driving, Lei Jun has stated that Xiaomi’s car project will not disclose any progress or respond to any rumors for the next two years.

All the suspense will be revealed when the delivery takes place in 2024.

— End —

This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.