Buick "Old Tree Blossoms New Flowers"

Author | Chai Xuchen

Editor | Wang Xiaojuan

In the wave of new energy, the once-dominant joint venture giant Buick, which has been somewhat quiet, is attempting to return to the center of the table through a radical technological bet.

After unveiling the Ultium platform 2.0 and the "True Dragon" range extender system, on August 18, SAIC-GM Buick officially announced a partnership with the autonomous driving unicorn Momenta to equip its first vehicle under the "Aito" brand, the L7, with the reinforcement learning-based autonomous driving large model—Momenta R6 Flywheel Model, which will debut full-scenario assisted driving features such as seamless city NOA and one-button parking without stopping.

It is understood that Momenta's R6 adopts a reinforcement learning large model, which differs from the previous "repeater" imitation learning large model; R6 is an evolutionary player. It can conduct massive simulation training in a virtual environment, iterating itself from both successful and failed attempts, exploring optimal solutions that surpass human drivers, and continuously optimizing through real road data feedback loops.

In a post-event interview, Momenta CEO Cao Xudong told Wall Street Insight that after equipping the R6, the Aito L7 will feature end-to-end AEB functionality, capable of braking instinctively like a human response without needing to detect an object first, even in extreme situations such as a slightly unexpected head turn.

Secondly, it offers a smooth experience akin to that of an "experienced driver." Based on the reinforcement learning large model, the Aito L7 can achieve truly seamless city NOA. Whether navigating narrow roads, making unprotected left turns, or autonomously passing through highway toll booths, the vehicle's predictions are more accurate, and acceleration and deceleration are smoother and more fluid.

With the technological support from Momenta, the Aito L7 undoubtedly gains a ticket to the first tier in terms of intelligence. This ticket is crucial for Buick, which seeks to reshape its high-end image.

Since entering the Chinese market, Buick has launched several flagship family sedans, from the early Buick Century and Park Avenue to the more recent LaCrosse Avenir. It is evident that Buick has been striving for high-end positioning, having once created a sales legend for joint venture cars, and many still have a Buick sentiment today.

As we enter the era of new energy, joint ventures have been pushed into a corner, but Buick has not sat idly by; it has launched a new high-end new energy sub-brand—Aito ELECTRA, with its first sedan, the L7, targeting the 300,000-level family car market. Buick aims to ride the wave of the times and regain lost years through sincerity and technology.

SAIC-GM General Motors Manager Lu Xiao emphasized that after launching the Buick high-end new energy sub-brand "Aito," six Aito models will be introduced to the market in the next 12 months. "The goal for these new energy products is definitely profitability, and they will all feature the next generation of smart cockpit and assisted driving technologies."

He set a new flag: by 2026, the sales proportion of SAIC-GM's new energy products will reach over 50%, and by 2027, it will even reach 60%, with all these new energy products needing to be profitable After regaining direction in the fog of chaos, Buick's battle for brand dignity has officially begun. From technology releases to product profitability, from strategic declarations to market realization, "speed" will be the key factor determining the success or failure of this gamble. For Buick, there is no turning back in this battle concerning brand dignity.

The following is a transcript of the dialogue between Wall Street Insights and SAIC General Motors General Manager Lu Xiao, Deputy General Manager Wang Chendong, Executive Deputy General Manager of Pan Asia Technical Automotive Center Zeng Yu, and Momenta CEO Cao Xudong:

Q: In the past two years, the speed of technological iteration in the intelligent driving field has been very fast. Many people feel that the technological innovation and evolution speed of intelligent driving companies is faster than that of OEMs. Is this view correct? In which technological fields are the two sides inseparable?

Zeng Yu: Decoupling is feasible. Each side has its own focus in the technological field: Momenta is professionally leading in intelligent large models; General Motors, as a century-old OEM, also has a deep professional accumulation and experience in vehicle performance, which is a complementary advantage for both sides.

Taking vehicle performance as an example, different vehicle types have different characteristics. SUVs or MPVs have a higher center of gravity than sedans, and if they turn at the same speed without targeted tuning, they will inevitably lean significantly. The differences in performance among different drivers on the same stretch of road also confirm the importance of tuning—some cause screams, while others are smooth and steady.

Our goal is to achieve excellent performance for each model through precise vehicle performance tuning, combined with Momenta's large model control, to realize perfect vehicle dynamics. This is the guarantee at the "muscle" level, relying on advanced electronic architecture. Its core function is to efficiently control all "muscles" and ensure that the "brain's" instructions can be accurately transmitted and executed.

This is precisely the value of Buick's new "Xiaoyao" super fusion architecture electronic architecture. The realization of assisted driving is far from the achievement of a single model; it requires the coordinated response of the entire vehicle system, including body structure, electronic architecture, intelligent cockpit, intelligent chassis, and even power systems. The key role of the electronic architecture is to coordinate these systems in a very short time and respond accurately to assisted driving commands.

Q: One of the hottest topics in the intelligent driving field right now is driving safety and reliability. How did both sides work together to develop a system with higher safety, and were there any disagreements or differences of opinion?

Cao Xudong: A strong alliance will definitely have collisions; only through collisions can we better understand each other's experiences and perspectives, creating better solutions.

General Motors places great importance on functional safety. For some potential safety risk scenarios, there are in-depth functional safety requirements and analyses. In addition to these analyses and decompositions, there are higher-level requirements for simulation and testing. It is not just theoretical analysis and decomposition; more importantly, there must be simulation validation and real vehicle validation. These validation test cases are built from the rich simulation testing scenarios created by the more than 300,000 vehicles we have already mass-produced and the data from 3 billion kilometers, and these testing scenarios are all derived from the feedback of real mass production data Another aspect is smoothness. During the development process, the management team at SAIC-GM has a habit of placing high demands on smoothness. A novice driver can accept a scenario where they brake to ensure safety, but an experienced driver insists on achieving extreme smoothness while ensuring safety, which presents us with very high requirements and challenges.

At the same time, the data from the test drive process, as well as the data from the "experienced drivers" at Pan-Asia, have been integrated into our data system and used as high-priority evaluation data in our development and testing systems to meet higher standards. Our friction and collisions are actually aimed at delivering better technology and a better product experience to consumers.

Q: Recently, the intelligent driving tests by Dongche Di showed that the performance of pure vision solutions is outstanding. Does this mean that pure vision solutions are the "better solution" for assisted driving technology perception?

Lu Xiao: Tesla's performance here is excellent, representing that the pure vision technology route is indeed worth pursuing and exploring further. However, technology routes can flourish in many ways, and our ultimate goal is to provide customers with a safer, more comfortable, and efficient intelligent experience.

Our collaboration with Momenta using LiDAR aims to cover a more complete range of scenario applications, including extreme road conditions. Combined with the support of the R6 flywheel large model algorithm, we can provide a more reassuring experience.

As mentioned earlier about smoothness, two weeks ago I led our management team to drive the Zhijing L7, which is equipped with Momenta's latest APA intelligent parking experience that you saw in the video, achieving true smoothness and providing the best experience for a 300,000-level intelligent luxury new energy sedan.

Wang Chendong: I experienced the latest version of FSD in North America at the end of last month, and its performance is indeed globally leading. However, this leadership may face adaptability challenges in the Chinese market. The assisted driving field should flourish in many ways, and there is no absolute optimal solution; the key lies in who best fits the local market.

The Chinese market has a significant long-tail effect. Relying on perception systems trained on American data, even with a large amount of data, if the sensor configuration is insufficient (such as with pure vision solutions), it may be difficult to fully guarantee the safety baseline we emphasize. Although data learning capabilities continue to improve, at this stage, to ensure safety, we insist on using LiDAR and other multi-sensor fusion solutions to guarantee safety through redundancy.

The future technology route will continue to evolve. Tesla may also adjust its strategy in the Chinese market. However, the parallel exploration of different technology routes is meaningful, and ultimately, they will learn from and integrate with each other to jointly promote the development of L3 and above high-level intelligent driving.

Cao Xudong: On the surface, it is a question of vision and laser, but at a deeper level, it is actually about the joint optimization of the "body" and the "brain." I would like to add another case: end-to-end AEB will also be equipped on the R6.

Previously, AEB involved first perceiving, then judging whether there was an emergency based on that perception, and then deciding to brake. The R6 has upgraded AEB to end-to-end AEB, which can brake without needing to detect an object, similar to a human's instinctive reaction This AEB is combined with a very good actuator from SAIC-GM, which shortens the execution delay and the model/brain delay, making the body very responsive and the brain also very sensitive, ultimately achieving smarter safety. As a new product highlight, it will be equipped on our R6 large model.

Q: What is the relationship between the collaboration with Momenta and SAIC-GM's SuperCruise and NOP? Will they share in the future or will they only use Momenta's solution?

Wang Chendong: We have three stages. Ten years ago, Cadillac's SuperCruise was very advanced globally, and we introduced it to China at the first opportunity. At that time, from its perception to functional safety to overall vehicle control, we had already considered everything very thoroughly from a system perspective.

Over the past time, on Cadillac, our Pan-Asia team has done a lot of work on how to better integrate chassis control, power system control, and steering control using algorithms and sensor inputs.

Of course, we are talking about the first generation here. Why is it very important to us? It's not just for the SuperCruise product; more importantly, after we gained this capability, we were able to develop our second generation—NOP highway navigation assistance system locally on the Buick GL8. Although the sensors and algorithms at that time were still different from today, from the start of NOP, the control was truly implemented entirely by our Pan-Asia team.

We have been emphasizing that strong partnerships actually have two aspects: on one hand, the perception layer and model algorithms, which we have now achieved industry leadership with the support of Momenta; on the other hand, our team's control of the entire vehicle, including functional safety settings and some scenario definitions, when combined, have achieved our product performance on the Buick Zhijing.

So in the future, this new driver assistance technology will certainly not be limited to Zhijing; there will also be opportunities to launch it on Buick models. Cadillac will also launch a driver assistance product based on Momenta in Q1 or Q2 next year.

Q: How do SAIC-GM and Momenta specifically divide their roles to create differentiated selling points for intelligent driving?

Cao Xudong: I think the key points are two: the first is insight into user needs and the pursuit of intelligent efficiency. When parking, if you need to step on the brake before entering parking mode, it will lead to two problems: the first is that the user's operational actions are excessive, and the second is that it increases the parking duration; if you don't need to stop, it only takes 18 seconds.

On the other hand, after identifying the need, can it be achieved? Not all OEMs can do this with us; it definitely requires the strongest "muscle" and the strongest "brain" to be closely integrated. The vehicle's actuators and our algorithms must be tightly integrated to achieve the ability to enter parking mode without stopping, smoothly linking driving and parking together for efficient parking. This also reflects the advantages we have gained from strong partnerships and deep integration Wang Chendong: Looking back at when we were defining hardware and vehicle controllers, this function actually did not exist; it was co-created later. During the implementation process, one challenge was to significantly challenge Momenta's algorithms, and the other was to challenge the control of the entire actuator. Both aspects are based on the hardware foundation we just mentioned, which does not have additional special sensors. We achieved this through software, algorithms, and actuator tuning.

Question: The current market is indeed very competitive. The ZhiJing L7 is a very important product for counterattacking the new energy market. What kind of differentiated competitive advantages will there be after the launch of the large model?

Wang Chendong: There are several exciting features that will be launched later, including stronger parking capabilities and how to implement more Door-to-Door functions in park scenarios. We are still focused on user needs, not just making a short video, creating an advertisement, or showing off; we hope to uncover the scenarios that users truly find challenging.

Cao Xudong: The first is the intelligent safety and peace of mind. Our reinforcement learning large model, compared to imitation-based large models, not only uses good human data but also utilizes challenging data or bad data from extreme scenarios. Our large model can explore good driving strategies in these scenarios.

This model does not learn human driving because human handling in these extreme scenarios or the original imitation learning handling is not very effective. However, through reinforcement learning, we can explore safer, more reassuring, and smoother driving strategies. Once this strategy is explored, the model will learn from it.

Just like AlphaGo initially imitated human experts and played chess with humans, at that time, it had a back-and-forth with humans. However, after significantly surpassing humans, all Go masters could not defeat AlphaGo because it used reinforcement learning. It even learned to play against itself, at which point experts evaluated that AlphaZero was already winning by 1 or even 2 stones against top human experts. In Go, a difference of 1 or 2 stones is basically a difference of 10 to 100 times.

Question: Are there plans for L3-level assisted driving in the future?

Wang Chendong: We have always been planning for L3. The redundancy we mentioned, especially for braking and steering, has redundancy in braking with three systems: electrical redundancy, mechanical redundancy, and electronic parking redundancy. Steering also has ECU redundancy and sensor redundancy. In fact, all these redundancies are prepared for L3.

In our planning, while doing L2, it is impossible to include all the hardware preparations for L3, but we have already considered it at the architectural, execution, and algorithmic levels. Last month, we held a board meeting in North America, where we reported the entire L3 strategic direction, path, and our product planning. Both shareholders recognized it very well. Our L3 is already on the road; it just depends on when the national regulations can be implemented, and subsequent models will be planned step by step Q: This time, why did Zhijing choose to use the Qualcomm 8775 chip, and what considerations led to not adopting an integrated cockpit design?

Wang Chendong: Indeed, the 8775 is the latest integrated cockpit chip, and many car manufacturers initially planned to use it in an integrated cockpit design. However, it also has its shortcomings, as it may not achieve optimal performance in single auxiliary driving scenarios. It can at most support high-speed navigation assistance or APA, but it has its limitations.

For us, we either do not engage in auxiliary driving, or we must do it well. Therefore, we are currently not considering using an integrated cockpit approach for the entire hardware; we will definitely use independent chips to support Momenta's latest and strongest algorithms.

Currently, we are using the 8775 in our intelligent cockpit because we value its AI capabilities, and the entire chip iteration is very fast. In terms of AI capabilities, it has a significant improvement over the 8295. In the future, our Zhijing L7 will have DMS, OMS, stronger voice recognition, and more microphones. We hope to leverage the hardware capabilities of this chip to ensure stronger AI capabilities and functional iterations in the cockpit domain.

Many cockpits are already very homogenized, as they all have similar functions and screens. However, I believe that AI, whether from the vehicle side or the cloud side, at least the 8775 chip can help us do more on the edge side