Last night in Beijing, there was another thunderstorm accompanied by gale-force winds. Even indoors, one could sense the severity of the weather, which is not suitable for driving as accidents are prone to occur. However, under such weather conditions, Ideal went out to test its own intelligent driving system.

01 Ideal's End-to-End Intelligent Driving Test in a Thunderstorm Night

From the video released by Li Xiang, the chairman of Ideal Auto, it can be seen that the rain outside was heavy, and the visibility was impaired, with the vehicles in front already using their hazard lights. It's important to note that in severe weather like heavy rain, the sensing ability of intelligent driving hardware such as cameras is greatly affected, which in turn impacts the performance of intelligent driving. However, Ideal's intelligent driving system performed steadily within the lane, accurately recognizing and responding at traffic lights, and even changing lanes accurately and quickly. Impressively, after turning right at an intersection, Ideal's intelligent driving system was able to recognize and navigate around an irregular obstacle. This performance is indeed commendable. Li Xiang revealed on his Weibo that this is breakthrough E2E+VLM technology for "autonomous driving." Before discussing this, let's first talk about Ideal's current intelligent driving system.

Currently, Ideal's intelligent driving system is divided into AD Pro and AD Max, with the former being the basic version and the latter being the advanced version. AD Pro has introduced Light Boat Navigation as an intelligent driving supplier, while Ideal's own efforts are focused on AD Max. In December 2023, Ideal pushed AD Max 3.0 to users. 3.0 models perception and control, while introducing BEV large models and Occupancy networks, enabling Ideal to achieve mapless NOA in cities.

Although AD Max 3.0 is already leaning towards end-to-end from an overall architecture perspective, it has not yet achieved complete end-to-end.

The intelligent driving version demonstrated this time can be considered a true end-to-end intelligent driving system.

02 Pushing L3 Autonomous Driving by the End of the Year

"We believe that the earliest by the end of this year, and the latest in the first half of next year, true L3 (supervised autonomous driving) can be delivered to users in batches," said Li Xiang at the 2024 China Chongqing Automotive Forum not long ago. Note that this refers to autonomous driving (L3 and above), not assisted driving (L0-L2.9). The "E2E+VLM technology" mentioned earlier is the confidence of Ideal in achieving autonomous driving. Ideal has built two systems for its autonomous driving system: System 1 (fast system) and System 2 (slow system).

This system concept comes from Daniel Kahneman's book "Thinking, Fast and Slow": System 1 (fast thinking): This is a fast, intuitive, and automatic thinking system. It hardly requires conscious effort and can quickly respond to information. System 1 handles daily tasks such as recognizing familiar faces, making simple decisions, and habitual behaviors. It relies on heuristics and mental sets, which can be fast but not always accurate. System 1 works well in many situations but may make mistakes in complex or analytically demanding situations. System 2 (slow thinking): This is a slow, logical, and effortful thinking system. It requires us to concentrate attention and energy on complex calculations, in-depth thinking, problem-solving, or evaluating the intuitive responses of System 1. System 2 is more reliable in handling tasks requiring logical reasoning, risk assessment, and long-term planning, but it is also more prone to fatigue and errors. In Ideal's view, end-to-end intelligent driving is data-driven, but there are countless unknown and long-tail scenarios in real life. Data-driven alone cannot fully solve the problems encountered in intelligent driving, and many scenarios require a true understanding and adaptability to real-world events.

Therefore, for the realization of L3 and L4 autonomous driving, Ideal has introduced a new approach: knowledge-driven. This gives rise to the two fast and slow systems. System 1 (fast system): performs rapid information input and response through end-to-end. System 2 (slow system): analyzes and reasons about the surrounding environment through cognitive models such as VLM and performs logical thinking. In subsequent decision-making, the results of the two decision-making systems can be fitted and compared to provide the optimal control solution.

Reflected in the vehicle, the intelligent driving system can be more human-like, with fewer false triggers. For example, in the future, when encountering vehicle advertisements on road signs while driving, the system will recognize them more accurately and will not falsely trigger AEB. Or, at complex intersections, it can make more human-like avoidance maneuvers. This year's upgraded Ideal AD Max 3.0 has replaced two Orin X chips with stronger computing power, so one is used to run end-to-end, while the other runs the compressed VLM model, which is just right. In addition, Ideal also has a cloud-based world model used to train these two fast and slow systems, thus forming a complete data closed loop. In the past few years, Ideal has also built a massive intelligent driving training cluster to achieve rapid iteration of intelligent driving model training (with a computing power of 1400 PFLOPS). For reference, XPeng's Fuyao Intelligence Center has a computing power of 600 PFLOPS; Haomo AI's MANA OASIS Intelligence Center has a computing power of 670 PFLOPS; Baidu has reserved a smart driving computing power of 1.8-2.2 EFLOPS for Jiyue; and Huawei's smart driving cloud training computing power is 3.5 EFLOPS (truly leading the way).

In the third quarter of this year, Ideal will officially roll out mapless NOA nationwide and will also push "autonomous driving" equipped with E2E+VLM technology to test users. Li Xiang said, "With the evolution of this technology, the enhancement of computing power, and the enlargement of the model, I believe that unsupervised L4 autonomous driving will also be achieved within at least three years." So, is Ideal's intelligent driving finally about to take off?

03 Is Ideal's Intelligent Driving Taking Off?

Ideal is now much more low-key than before, focusing on technology development. I know some people may be overjoyed to see Ideal's performance, but I have to pour cold water: it's still just an internal test, and there's still a long way to go before official rollout.

For now, Ideal's intelligent driving is quite strong, but there is still a gap compared to absolute leaders like XPeng and Huawei. It should be noted that XPeng and Huawei are moving faster in terms of end-to-end intelligent driving in China.

XPeng pushed the first domestic end-to-end intelligent driving system in May, and by 2025, XPeng vehicles will achieve L4-like intelligent driving experiences in China. Huawei is likely to push ADS 3.0 for end-to-end intelligent driving in August this year, with the first NCA (Navigation on Autopilot) from parking space to parking space. So, can the new intelligent driving system using "E2E+VLM technology" help Ideal regain the upper hand? I can't say for sure. But what I know is that Ideal is holding back now and will burst out when the time is right. End.