Li Auto's brand new generation of driving technology architecture end-to-end + VLM system officially fully pushed.

LI Auto officially fully push the new generation dual-system intelligent driving technology architecture end-to-end + VLM, marking the official entry of LI Auto's intelligent driving into the era of AI large models, truly achieving human-like thinking and driving like an experienced driver, full coverage of scenarios, more comfortable driving, and more efficient passing.

Based on LI Auto's continuous emphasis on safety, it has been continuously refining and verifying the dual-system architecture. After multiple rounds of testing involving thousands to tens of thousands of people, LI Auto's intelligent driving training mileage has now exceeded 2.5 billion kilometres, with NOA mileage accounting for 80%. The proportion of AD Max orders for models above RMB 300,000 exceeds 70%, with L9's AD Max orders exceeding 80%. The training computing power has reached 0.539 billion Eflops. In addition, thanks to LI Auto's high-quality training data and high computing power, as well as the advanced deployment of the world model in the cloud, it can achieve efficient iteration of two versions per week, leading the industry. This full version is based on the latest V4.8.6 model, with increased urban scene data ratios, enhancing overtaking scene understanding, navigation information understanding, obstacle detection accuracy, stop line detection capabilities.

OTA 6.4 version's intelligent space adds Task Master 2.0, fully integrating LI Auto's classmates and Mind GPT capabilities. With the support of large models, Task Master becomes smarter and more comprehensive. The navigation map is upgraded to the Gaode AutoSDK 750 version, further enhancing the functionality experience. The smart electric vehicle adds ideal 4C/2C/preferred station supercharge report and charging station power range filter, efficient and stable, worry-free charging in urban areas and highways.

On Oct 23, 2024, LI Auto Inc, the pioneer in the ideal car industry, officially fully pushes the new generation dual-system intelligent driving solution end-to-end + VLM (Vision Language Model) to LI L-Series AD Max users and LI MEGA users with the OTA 6.4 version of the vehicle system, marking the official entry of Li Auto's intelligent driving into the era of AI large models. This OTA update focuses on upgrading intelligent driving capabilities, with intelligent space and smart electric vehicles also updated simultaneously, providing LI L-Series and LI MEGA users with a more comprehensive intelligent travel experience.

The end-to-end model has reached a new level of humanization, with driving habits comparable to experienced drivers.

Li Auto 's end-to-end + VLM integrates a true One-Model structure end-to-end model and the globally innovative VLM visual language model, and is the first to deploy the VLM visual language model on the in-car chip dual-system solution. From the stage of innovative technology to actual implementation, the level of intelligence, humanization, and stability continues to steadily improve.

In this OTA 6.4 full version, in terms of end-to-end capabilities, new functions such as roadside start, U-turn, and roundabout passage have been added, enhancing lane change, detour, and yield in congested road conditions; precise lane selection at complex intersections; and recognition and response capabilities for scenarios involving vehicles or pedestrians merging and crossing. Truly achieving seamless urban scene transitions, easy passage through complex roundabouts, smooth U-turns, more confident congested driving strategies, enabling Park gear activation in situ, providing users with a more convenient and enjoyable smart driving experience.

When in a roadside starting scenario, the end-to-end model will directly output trajectories, using stronger feasible area planning, no longer heavily relying on lane relationships. Even if the vehicle is not in a lane, it can heuristically find a route from the roadside to the main road, allowing for a direct start with a double click of the lever in Park gear, without the need to start in the lane. Thanks to the end-to-end model's powerful perception abilities, it can more accurately identify and predict the driving trajectories of surrounding obstacles, ensuring the safety of roadside starts, and during the starting process, looking around to determine dangers, autonomously yielding and giving way.

The end-to-end model is trained based on human driver data, possessing stronger understanding capabilities for complex road structures. For example, in different U-turn scenarios, it can output reasonable trajectories without heavily relying on navigation information, road topology, and manually defined rules to choose the appropriate route. Stronger lateral perception abilities achieve "seeing far, seeing accurately", during U-turns, it can more accurately identify and predict the driving trajectories of oncoming straight vehicles and take evasive actions, in roundabout passages, it can more accurately identify and predict the driving trajectories of merging vehicles to engage in reasonable gameplay, making more decisive decisions entering and exiting roundabouts and choosing exits more precisely.

In addition, end-to-end, through learning a large amount of data from experienced drivers in various driving scenarios, relying on pure model-based algorithms, achieves more human-like lane change and creeping actions. In the event of the preceding vehicle decelerating, it will directly bypass creeping through; in complex intersections with unprotected left turns, accurately assesses traffic gaps and proactively selects lanes with less traffic flow. Due to the lower end-to-end latency from perception to control, it has the ability to react quickly and decelerate promptly when encountering situations of vehicles or pedestrians cutting in or crossing, balancing safety and traffic efficiency.

VLM visual language model ensures compliance with traffic regulations for more reasonable and efficient traffic flow.

As of now, Li Auto's VLM visual language model deployed at the edge has 2.2 billion parameters, possessing a more human-like understanding of the complex traffic environment in the physical world. It can adeptly handle unknown scenes even for the first time. With the official push of OTA 6.4 version, VLM visual language model has also received significant updates, learning from a large amount of human expert drivers' driving data, driving behaviours in special lane scenarios, and challenging driving situations to comply with traffic regulations, recognize Chinese information on roads profoundly, identify tidal lanes, and traffic conditions, assist in completing the entire driving cycle end-to-end.

For example, in a timed bus lane, it can easily understand the information text of the overhead road signs and recognize the yellow markings of the bus lane, and correspondingly make control execution; can identify tidal lanes and variable lanes, as well as the signs of passable and no-entry states, and correctly choose lane passage; in T-shaped intersection scenarios, it will actively take defensive deceleration to avoid the feeling of insecurity caused by entering the bend too fast in the past. It can identify speed bumps, potholes, dark roads at night, school road sections, construction scenes, main and auxiliary road exits and entrances, and decelerate in advance, while providing warnings through voice/EID/UI interactions; in main and auxiliary road/elevated scenes, it can identify and understand the self-position of the vehicle as the main road/auxiliary road, elevated up/down, correct navigation deviations, ensuring the accuracy of route selection.

Furthermore, the OTA 6.4 version also optimizes the high-speed NOA function. In high-speed & urban expressway scenarios, for slow-moving vehicles ahead, it has the capability to identify and initiate overtaking actions earlier and decisively, making it more efficient and safer.

The dual-system interactive experience has been completely upgraded, providing a more convenient user experience.

In addition to the significant upgrade of the dual-system in intelligent driving functions, the user interaction experience has also been enhanced. After the full version push of OTA 6.4, li auto inc's end-to-end + VLM intelligent driving dual system display on the vehicle-mounted interactive page is clearer and more comprehensive, enabling real-time interaction of "seeing and thinking," making the entire intelligent driving experience more understandable and reassuring.

System 1's "Quick Thinking" end-to-end model text pop-up is more rich and easier to understand compared to the imageless NOA version, with real-time interaction according to the logic and actions related to navigation, traffic rules, efficiency, games, attention reminders, etc. System 2's "Slow Thinking" VLM visual language model adds graphic and text windows. In special scenarios, the images perceived in the front are projected onto the page, combined with text explanations of the model's thinking process and results. This allows users to "see and feel" the intelligent and convenient aspects of the dual-system intelligent driving functions during actual vehicle use.

In terms of intelligent space, Li Auto Inc always focuses on user value, with user experience as the core, leveraging AI artificial intelligence to create a benchmark intelligent cabin experience. In OTA 6.4 version, the Task Master 2.0 is added, fully integrating the capabilities of Li Auto classmates and Mind GPT, making the Task Master more intelligent and comprehensive with the support of a large model.

Navigation map upgraded to the high AmapAutoSDK 750 version, adding green wave speed, intersection meeting warning, green light countdown functions, enhancing the map's 3D visual effects, making the map more powerful. In addition, when initiating an along-the-route search in the navigation, specific keyword searches can be specified through touch control plus voice input. Fine filtering is available for specific types of along-the-route searches, for example, when searching for charging stations, brands and power can be filtered, and for gas stations, brands can be filtered. Travel-related Q&A results will generate voice map cards in the Li Auto classmates' voice box, combining text and images to show more vividly. Additionally, there are new desktop card speed display intervals and intelligent commuting card switches.

Ideal classmate adds emojis and eye tracking, adds thinking, happy, and puzzled expressions, under the array microphone blessing, Ideal classmate's eyes can follow the speaker's position; Ideal classmate's voice model is also upgraded, offline voice approaching online effects, English broadcast more natural; Ideal classmate's car control response speed is further improved by 10%-20%; the entertainment assistant adds entertainment application content on-demand semantic direct access; the car assistant adds OTA-related Q&A and operation instructions.

MEGA model adds a cabin rear view application to check the situation of rear children/passengers at any time; MEGA model adds a second-row right one-click luxury seat, one-click to enjoy the ultimate riding experience; L series adds one-to-three-row intercoms, easy conversation even in high-noise environments.

Regarding intelligent electric vehicles, with more and more ideal supercharging stations, supercharging reports naturally cannot be left behind. Added Ideal 4C/2C/Preferred Station supercharging reports. After charging is complete, a supercharging report is automatically generated, and both pure electric and extended-range models can view the details of this charge, such as duration, mileage, peak power, average power, etc., at a glance. Can be sent to the phone with one click, viewing, saving, and sharing seamlessly;

Also, a charging station power range filter is added, making filtering adjustments more intuitive, making it easy to find suitable charging stations, efficient and stable, realizing worry-free urban and highway charging.

Having 1 million vehicles is a very crucial indicator. Recently, Li Auto has completed the delivery of the first million vehicles. At the same time, thanks to Li Auto Inc.'s long-term adherence to core self-developed technologies and continuous increase in R&D investment, this time, Li Auto Inc. has launched the industry's first innovative new generation of fully integrated smart driving solution - end-to-end + VLM, also marking a significant milestone in full push. From now on, Li Auto's smart driving officially enters the AI big model era, bringing a more comfortable, comprehensive, and efficient smart driving experience to millions of Li Auto Inc. users. In the future, Li Auto Inc. will continue to make comprehensive efforts in smart driving, smart electric power, smart space, aiming to become a global leading ai company, and continuously striving to create a mobile home and a happy home for more users.