Author: The Office
Li Xiang has been active on Weibo recently.
On the afternoon of January 11, 2022, Li Xiang saw a video link from Donchengdi in an active safety-themed Feishu group within Ideal Automotive. Despite the video’s lengthy title, “Annual Selection of New Energy, AEB, Performance Models and Engines in 2021,” for Li Xiang, the focus was on three letters: AEB.
AEB, short for Autonomous Emergency Braking, is an automatic emergency braking system. During vehicle operation, if there is a risk of rear-end collision and the driver does not brake or applies insufficient force, AEB will be activated to brake strongly and avoid rear-end collision.
In 2021, Donchengdi tested AEB in over 100 cars with a variety of tests, ranging from child pedestrian detection and tunnel AEB to disappearance of the preceding vehicle and 50% rear-end collision. These tests went beyond the “questions” posed by collision safety evaluation agencies worldwide, and subjected AEB to rigorous testing that suited China’s traffic scenarios.
The video’s conclusion was that the 2021 Ideal ONE beat the top five contenders: Huachen BMW 5 Series, Huachen BMW X3, Subaru Outback, and Volvo XC60, and became the best-performing model in Donchengdi’s 2021 AEB test.
A few hours later, Li Xiang expressed his appreciation for the intelligent driving team on Weibo, saying, “Since we have been able to carry out full-stack R&D, the team’s progress in algorithms and control has far exceeded my imagination.”
For Li Xiang, and even Ideal Automotive, this was undoubtedly a small victory. Especially considering that just a year ago, at the end of December 2020, Ideal had just recently parted ways with the vision perception supplier Mobileye, and was already starting from scratch.
Why Ideal?
Everything goes back to the end of that year, over a year ago. Ideal had been cooperating with Mobileye for more than a year, and the collaboration had driven Ideal to the brink. This vision perception supplier, starting from Israel, had knocked on the doors of many mainstream automakers with its efficient and reliable algorithms, including the young Ideal Automotive. However, due to its highly closed cooperation strategy, Ideal was not allowed to continuously iterate and develop based on Mobileye’s algorithm foundation to fix some bugs collected from user feedback.
As Li Xiang said, “Since we cannot meet the demand of our self-developed full-stack intelligent driving (and the most crucial perception algorithm is a black box), we stopped cooperating with Mobileye at the end of 2020.”
It’s not difficult to realize the need for “full-stack self-development,” but in terms of timing, Ideal was quite passive.One industry norm for reference is that the two pioneers in self-developed intelligent driving, Tesla and XPeng Motors, both experienced over 6 months of R&D period for their first self-developed intelligent driving models (Model S in October 2016 and P7 in July 2020) before pushing out active safety features like AEB.
Considering that the SOP of the new Ideal ONE is scheduled for May 2021, even if the team and organization are in perfect condition, the time left for the Ideal intelligent driving team has become extremely limited.
However, Li Xiang’s requirement is that the 2021 Ideal ONE cannot be delivered without AEB. In the impossible triangle of time, R&D and goals, the 2021 Ideal ONE was eventually delivered with the basic AEB function equipped with Bosch millimeter-wave radar. This is what Li Xiang mentioned on Weibo as “basic AEB function delivered in June 2021”.
But in the context of full-stack self-development, this transitional solution cannot make the Ideal intelligent driving team more at ease.
In May 2021, Ideal Motors released the 2021 Ideal ONE at its headquarters in Shunyi, Beijing. At about the same time, a 5-person AEB self-development team was formally established.
Since there is an opportunity to start from scratch without relying on suppliers, “what to do” has become the first key philosophical question. If it’s just indiscriminately reinventing the wheel, why bother with self-development?
Ideal sorted out the current AEB market landscape. For a long time, the AEB market in China has been dominated by Tier 1 suppliers such as Bosch, Continental, and Autoliv, as well as Tier 2 suppliers such as Mobileye.
The research and development of these multinational suppliers is mostly carried out at their headquarters, and the main scenarios developed are mostly based on typical high-frequency accident scenarios abroad. The performance evaluation of AEB is also mainly based on the “test questions” of the European New Car Assessment Programme (E-NCAP).
The biggest problem with AEB developed by these multinational suppliers is that they have not developed scenarios for China, such as elderly mobility scooters with unclear right of way roaming on Beijing streets, food delivery motorcycles violating traffic rules and reversing on Shanghai streets at any time, and free-flowing traffic on non-motor vehicle lanes in Guangzhou.
An engineer discovered that for the Chinese market, the AEB test by the reputable collision institution E-NCAP is not even as valuable as some media platform AEB tests.
After sorting out, the 5-person team established the R&D goal of Ideal AEB: to develop Chinese-specific AEB based on Chinese characteristics of traffic scenarios and high-frequency accident scenarios while meeting the collision test standards of institutions.The core issue after defining what to do is naturally “how to do it.”
Despite the presence of multinational giants with R&D expertise exceeding 10 years, the ideal AEB self-developed efficiency has demonstrated its sharpness upon its debut.
From an external perspective, the ideal AEB’s R&D efficiency has been impressive, with the project starting in May 2021 and being launched in December 2021. However, from an internal perspective, these 6 months can be broken down into distinctly different time periods, each with its own story.
Starting from the project approval, less than two months were needed for the ideal AEB’s first feature, on-board activation, to be implemented. Of course, the performance at the time was self-evident.
Following the activation, a large-scale performance calibration was conducted, spanning various testing grounds from Chongqing to Yancheng in Jiangsu. With a focus on covering intensive and diverse traffic scenarios typical of China, as well as the basic traffic scenarios, Ideal made significant progress in the forward performance calibration of AEB by the end of July.
Compared to multinational suppliers whose R&D cycles are typically measured in quarters, this sounds suspiciously fast. However, if we look deeper into the R&D process, the secret lies in what was earlier referred to as the “algorithm black box.”
What is the “algorithm black box”? Perhaps we need to talk about AEB itself.
From the perspective of R&D, AEB can be roughly divided into perception and control. Perception acts like the human eye, responsible for receiving inputs from cameras and millimeter-wave radars and playing the role of “seeing.” Control acts like the human brain, responsible for issuing corresponding instructions based on the perception signals, such as not braking, braking with a certain force, or slamming on the brakes, playing the role of “thinking and decision-making.”
The difficulty of AEB lies in the fact that, in the public’s perception, this function is inconspicuous and is the most basic function in assisted driving. However, from another perspective, AEB is the only assisted driving function that requires waiting at all times in all scenarios because it is strongly related to collision risk. If there is a very low probability of AEB not performing when it should or of triggering when it shouldn’t, AEB will bring about enormous safety risks.
This makes multinational suppliers proceed with caution when iterating the development of AEB.
For example, when these multinational suppliers deliver their client vehicles in China, the owner of the vehicle may discover and provide feedback on a problem with AEB, such as late brake actuation, where the system identifies the risk of a collision in advance but with a delayed response. This indicates that the strategy of the control module needs to be adjusted.
For multinational suppliers, a demand like “adjusting one parameter of the control module,” will go through several layers of reporting to the global headquarters in Stuttgart, Germany, Dublin, Ireland, or Jerusalem, Israel. The request will then go through a series of validation calibration, testing, and risk assessment, until final approval. The average iteration cycle is at least 2-3 months.
You may ask, can perception be faster?
The perception scenario is more complex. The application scenarios of AEB are 24/7 in all scenes, which makes the demand for perception no different from that of L4 level fully autonomous driving systems.
Under such pressure, every multinational supplier will treat the iteration of the perception module with the most cautious attitude, and use a headquarters platform-based perception module to adapt to markets all over the world. Unless there is a security risk that has fermented on a massive scale, most feedback on perception problems in regional markets will be suppressed locally and will not be easily iterated.
This involves the cost considerations of adjusting the global market with a slightest change, as well as the time cost considerations for traditional suppliers based on their own vehicle fleet to complete a series of processes such as data collection, labeling, training, and verification. This also partly explains why Mobileye, the visual perception leader, insists on refusing to open its algorithms to automakers – the algorithm know-how accumulated by Mobileye for more than ten years is its most valuable asset.
In short, the contradiction between the AEB development process under traditional production relations and the high-frequency iteration demand of intelligent vehicles in the age of intelligence is basically irreconcilable. In most cases, when the whole vehicle project development reaches the SOP stage, the AEB project will also enter a short maintenance stage and will soon be completely terminated. “Iterative updates throughout the vehicle’s life cycle” is nothing more than a fantasy.
However, ideal cars, or more precisely, automakers with full-stack self-developed capabilities, because they have perception and control in their own hands, can return to a forward-looking logic of problem-solving without considering whether it is perception or control, locate research and development based on bugs, and call on other teams’ support to fix bugs at any time.
When doing performance calibration in Yancheng, Ideals also encountered the problem of modifying a parameter in the rule control module. However, because there were no departmental walls between different teams, they were able to iterate algorithms at a daily rate, so-called “solving the problem the same day it arises”. Even the perception module, which the supplier “doesn’t want to touch,” continues to be iterated on a monthly basis.
Opening the “black box”, connecting perception and control, this is the secret of ideal AEB’s rapid progress. However, this is only one of the secrets, let’s continue.
Since August, Ideal has started the robustness verification of AEB. Robustness originally means the ability of computer software to survive in exceptional and dangerous situations. In the context of Ideal, it means whether this new-born AEB, which has performed exceptionally well in the test environment, will still perform well when it is put on the open road?
The AEB team first carried out a round of 100,000-kilometer data collection through Ideal’s internal test fleet. For AEB, this is a mileage threshold with statistical significance. The mileage scale of AEB real-car verification in the industry is mostly between 100,000 kilometers and 200,000 kilometers.The reason for conducting such large-scale real-car verification is because good forward performance often only represents the A-side performance of AEB, and when we want to comprehensively evaluate the performance of AEB, the frequency indicator of B-side false triggers is indispensable. Blindly improving forward performance often means that the false trigger rate will also quickly increase.
The industry usually considers 1 time/100,000 kilometers or 1 time/160,000 kilometers (corresponding to 100,000 miles) as a basic threshold for acceptable false trigger rate. In other words, if you drive about 20,000 kilometers a year, encountering a false trigger of AEB once every 5 years on average is considered an acceptable probability.
By mid-August, Idealsec completed the 100,000-kilometer level test and repaired the false trigger bugs exposed.
However, the AEB team realized that such tests may conceal some issues. First, from the background data of Idealsec, the daily driving total mileage of the delivered Idealsec ONE fleet is about 3.5 million kilometers, which is insufficient in mileage and limited in sample size compared with the real-car verification of AEB. Second, compared with user scenarios from all over the country, the road test scenarios of Idealsec’s internal fleet are not enough.
They came up with the idea of “shadow mode”. From the post-results, “shadow mode” indeed showed great efficiency advantages and is undoubtedly the second secret weapon of Idealsec’s self-developed AEB.
This is a development method first proposed by Tesla. By reserving the corresponding software environment and computing power in the early stage of development, immature algorithms can continue to run in the background like shadows without intervening in the control of the vehicle. When the human driver controlling the vehicle and the algorithm in the background make different decisions, the system will record the corresponding scenes and data and upload them to the cloud for research and development analysis.
Idealsec’s head of intelligent driving, Lang Xianpeng, recalled that as early as April 2018, Idealsec clarified that the future of autonomous driving must be data-driven. Therefore, even though the 2020 version of Idealsec ONE is equipped with the Mobileye perception black box that cannot be updated, Idealsec still reserved an independent camera and corresponding software environment on the front windshield for subsequent data collection through shadows and the iteration of algorithms.
This time, Idealsec’s “shadow mode” was used for the first time in independent function development. On August 29, Idealsec first pushed its self-developed AEB algorithm to more than 100 early-adopter users’ vehicle models equipped with internal test versions in the form of a shadow mode.
The previous concerns were ultimately proven to have some foresight. From the data returned from the shadow mode of the internal test version, the catastrophic false trigger rate of the first internal test version of AEB reached 17-18 times/100,000 kilometers.Amidst the cold sweat, the AEB team also realized the great value of shadow mode to reduce false triggers and improve algorithm robustness. As the algorithm runs in shadow mode, signal collection and scene data feedback are silently deployed and completed in the background, with no impact on the user experience. The ideal thus quickly expanded the deployment scale of the AEB algorithm under shadow mode.
After the initial 20 million kilometers of “user road testing”, a large number of high-value false trigger scenarios were returned to the Ideal R&D team. Ideal then combed through all the triggered scenarios, split different bugs into different business modules, and brought together perception and control to fix them. The next step after bug fixing is called “simulation”. By simulating similar scenarios in a virtual environment, Ideal verifies the robustness of new algorithms.
Like the initial R&D, Ideal also brought together perception and control modules for joint simulation in the new algorithm’s simulation phase. Unlike the traditional supplier’s approach of conducting perception simulation and control simulation separately, this “one-track” mode requires perception simulation to be completed before running control simulation, catching many bugs that were previously missed in the simulation process.
Afterwards, Ideal iterated on two versions of the AEB algorithm on a monthly basis through shadow mode. Data shows that the final AEB false trigger rate has been reduced to 0.7 times/100,000 kilometers, equivalent to a maximum of one AEB false trigger encountered by an Ideal car owner during the entire lifespan of their vehicle.
Regardless of subsequent continuous iterations, even up to this point, the development of Ideal’s AEB has already experienced over 100 million kilometers of official road test data, a three-order of magnitude reduction in dimensionality compared to the traditional supplier’s 100,000-kilometer level.
On December 7th, the first mass-produced version of Ideal’s self-developed AEB was officially released.
Moreover, this version defeated established players in the active safety field, including BMW, Subaru, and Volvo, in the annual test conducted by the well-known Chinese automotive media, D车评.
Looking back at this, it is still somewhat unbelievable: a new brand that self-developed AEB from scratch, with a team of less than 10 people, defeated multinational giants with more than 10 years of accumulated experience in AEB development in just seven months.
Fundamentally, the production relations and production materials of AEB R&D have undergone a significant change in Ideal.
Before 2016, multinational automotive brands across the world had never considered personally engaging in the self-development of AEB. Tesla’s appearance showed everyone that self-development not only is feasible, but also brings a revolution in experience and efficiency.
We can’t help but ask: how did the notion that AEB R&D has an extremely high threshold come about?Before car manufacturers began to manage customer relationships on their own, there was a natural gap between car manufacturers and customers.
In an era without whole-vehicle OTA and user data to support iteration, the complete development, performance calibration, and false-trigger rate testing of an AEB system could only be conducted by the supplier’s internal car team, resulting in high costs in manpower, finance, and time.
The complete iterative development of the perception and control chain and zero-cost “shadow mode” data input has formed a significant advantage over traditional AEB development mode.
We must accept the truth that the core competitiveness of traditional AEB does not lie in its performance and false-trigger rate in actual driving scenarios, but in the high development costs that create a competitive barrier. When “data-driven” advances on both cost and efficiency, the traditional AEB with a low actual performance limit becomes an empty tiger.
What’s more, AEB’s evolution has just begun. With user data support, the AEB team of Li ONE, with the help of Ideals Automotive, collated all the recorded cases of accidents that had occurred on the delivered Li ONEs and classified them, and then separated the information to various sections of Ideals Intelligent Driving, providing guidance for the continuous iteration of AEB.
At this stage, AEB research and development has broken free from E-NCAP testing and professional media scenarios, and completely entered the user’s driving scenario, precisely attacking potential safety hazards, and this is the real forward-looking development.
In May 2021, Ideals announced the 2021 model of Li ONE at its headquarters in Shunyi, Beijing. Six months later, this car surpassed Toyota Highlanders that had individually dominated the large and medium-sized SUV market for a long time, with a monthly sales volume of 13,485, becoming the new sales champion. In the perception of many, Li ONE stood out with its excellent product strength.
However, looking at AEB self-research strategy glimpses the overall situation, and Ideals Automotive is rapidly advancing its technical self-research strategy. As Li Xiang said, investing in technology research and development, not just in product research and development, is a fundamental turning point for Ideals Automotive after its IPO.
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.