It has been less than two months since the Chengdu Auto Show announced the recruitment for a large-scale experience of the brand-new generation of intelligent driving systems from Li Auto. Today, the company officially rolled out their latest "End-to-End + VLM" dual-system intelligent driving solution. This intelligent driving system stands out by enabling vehicles to think and drive like humans, achieving full coverage of scenarios, providing a more comfortable driving experience, and ensuring more efficient passage.
Li Auto's industry-first "End-to-End + VLM" full rollout
Simultaneously, Li Auto's vehicle-machine interaction for intelligent driving has also been updated. The "End-to-End + VLM" dual system provides a clearer and more comprehensive display on the vehicle interface, giving users peace of mind throughout the intelligent driving process.
As competition intensifies in the realm of intelligence, Li Auto has made remarkable strides. Reflecting on its journey, the company has transformed from a latecomer in intelligent driving to leading the first tier. From Highway NOA to urban uncharted NOA, and now to the complete rollout of the End-to-End + VLM model, this technology framework operates much like an "experienced driver," significantly enhancing safety, comfort, and efficiency compared to the uncharted NOA.
Having reached this point, Li Auto is no longer using traditional methods for autonomous driving but is, for the first time, implementing artificial intelligence in this domain. This shift reflects the company's contemplation of user value and its understanding of autonomous driving.
We hope to provide users with the feeling of an experienced driver helping them behind the wheel.
The "End-to-End + VLM" signifies Li Auto's official entry into the era of large AI models in intelligent driving, serving as a watershed for differentiating intelligent driving technologies. The term "end-to-end" refers to a development paradigm involving a task that spans from the initial input to the final output without intermediate processes. Li emphasizes that their end-to-end model is genuinely representative. "We believe that end-to-end is an extremely efficient model that allows for rapid iteration and execution," said Lang Xianpeng, Vice President of Intelligent Driving at Li Auto.
End-to-End - Fast Thinking Visualization
VLM can be understood as a large language model tailored for autonomous driving. It equips the system with logical thinking capabilities to handle unknown scenarios, enabling it to deeply understand the complex traffic conditions of the physical world and the nuances of Chinese semantics. Furthermore, it supports the end-to-end model in tackling industry challenges related to safety, navigation, regulations, and comfort.
Additionally, Li Auto has independently developed a reconstruction and generative world model for the training and validation of their intelligent driving solutions.
This represents Li's latest approach to autonomous driving. "If we truly want to achieve complete autonomous driving—where the system takes over completely—it must understand physical events as a human would," Lang Xianpeng explained. "Using maps or lacking them won't accomplish this, as map limitations will always prevail."
"Ordinary consumers don't really care whether the system is map-based or not, whether it’s end-to-end or not. What they care about is the product and the user experience—essentially, the product's value," Lang added. "We hope to provide our users with the sensation of an experienced driver helping them behind the wheel."
With the "End-to-End + VLM" launch, Li Auto's vehicles equipped with advanced intelligent driving functions can now tackle various challenging everyday driving scenarios, such as starting from a stationary position by the roadside (activation from park mode without lane initiation), navigating roundabouts (decisively choosing exits while entering and exiting), and making U-turns (suitable for more cases, including scenarios where there's a flower bed by the roadside).
Moreover, leveraging Li’s unique VLM, Li Auto can now recognize and respond to scenarios such as limited-time bus lanes, tidal/variable lanes, speed bumps, and uneven road surfaces.
Enhanced Decision-Making in Circles with Reasonable Strategy
Furthermore, Li Auto's Highway NOA has seen significant optimization, enhancing its ability to identify slower vehicles early and execute overtakes decisively in high-speed and urban rapid transit scenarios, making driving more efficient and safer.
"Many people used to buy Li vehicles for space, large displays, and spacious sofas, but in the future, they may purchase the car primarily for intelligent driving," Lang Xianpeng emphasized. "Any product Li develops must exceed or meet user value."
In the future, vehicle-centered autonomous driving will focus on data and computing power as core competitive advantages.
Li Auto's new generation "End-to-End + VLM" dual-system intelligent driving solution is not only a world premiere but has also been completely rolled out. The company’s rapid progress in intelligent driving is partly attributed to organizational efficiency, as the R&D team keeps pace with the latest technological advancements in intelligent driving. Additionally, Li Auto employs a tiered R&D process, allowing the team to simultaneously conduct preliminary research on end-to-end while developing uncharted NOA, balancing efficiency and quality.
As mentioned earlier, end-to-end represents a watershed for distinguishing intelligent driving technologies. Why is it a watershed? Beginning with end-to-end, Li Auto has shifted away from traditional automatic driving design thoughts, with data becoming the true driving force. "From our own end-to-end model perspective, we just need to treat it like a 'seasoned driver,' aiming to deliver the same driving experience, using driving data from all experienced Li vehicle owners," Lang explained. "Li is very strict in filtering this data; out of 800,000 owners, only 3% truly represent the data of seasoned drivers."
The VLM has 2.2 billion parameters and exhibits highly human-like comprehension abilities.
Li's sales figures and the scale of its owners are extraordinary in expanding the limits of their autonomous driving capabilities. The data used for model training comes from here, and having more and better data, coupled with the necessary computing power for model training, will be central to future competitive advantages. "Everyone is trying to reach a visible ceiling, but the challenge lies in how to enable the system to feel more human-like and accelerate iteration. At this stage, companies without sufficient data and computing power will be the first to be eliminated," Lang warned.
To excel in the future competition of autonomous driving, companies not only need ample data and computing power but also substantial funding. In the second quarter of 2024, Li Auto's R&D expenses reached 3 billion yuan. The company allocates 10% of its revenue to R&D in core technologies such as intelligent driving, smart space, smart electrification, and vehicle safety. Li estimates it will spend 1 billion dollars annually on computing power training in the coming years. Lang disclosed, "This figure does not include various personnel expenses; companies without a net profit of at least 1 billion dollars cannot afford this investment."
Intelligent driving is gradually becoming a significant consideration for the new generation of young car buyers. To date, Li Auto has achieved over 99% user penetration for its intelligent driving technology, with the total distance driven exceeding 2.5 billion kilometers—83% of which is attributed to NOA miles. The total mileage for all scenarios of NOA has exceeded 1.35 billion kilometers; in models priced over 300,000, the order proportion for AD Max has reached 70%, particularly for the L9 model where it stands close to 80%.
The latest generation V4.8.6 model genuinely enables thinking like a human and driving like an experienced driver.
"As we supplement and iterate on data and computing power, the emergence of urban autonomous driving derived from the end-to-end architecture holds great promise for replicating the experiences found on highways. If we can achieve such experiences, I believe we've crossed a significant threshold," Lang noted. He believes that once we reach the stage of End-to-End + VLM, the vehicle will take the lead, with humans merely playing a supportive supervisory role. "Achieving this level will fulfill our users' needs for autonomous driving, and crossing that threshold will undoubtedly greatly motivate car purchases."
Currently, Li Auto is the only new car manufacturer in China to achieve sales of over one million units, consistently selling over 50,000 units monthly.
In Closing
With multiple generations of intelligent driving products, Li Auto continues to make new breakthroughs in this field while providing the industry with a new technical paradigm. "We iterate on technology not for the sake of technology itself but because this technology genuinely meets user needs and enhances product experience," Lang remarked.
Li Auto's next step is to continuously invest in R&D in the field of artificial intelligence, aiming to use AI-driven machines to transform the efficiency and experience of the physical world, benefiting every family served by Li Auto and every member within those families.
