The pure vision technology route appears to be cooling down due to an unexpected incident, prompting the industry to recognize the advantages of the fusion perception route in terms of safety and reliability. This debate over technology routes persists.

As autonomous driving technology suddenly faces stricter industry regulations, automakers are shifting their focus from promoting intelligent vehicle capabilities to ensuring safety. The pure vision technology route seems to be quietly waning. However, just during the May Day holiday, a message released by Tesla reignited heated discussions within the industry.

On May 4, Tesla's official Weibo account posted a promotional image with the text, "Adhering to the visual processing solution, making safe and intelligent products affordable for everyone." The accompanying text emphasized, "Tesla has proven with its strength that advanced technology does not require expensive and complicated sensors." Interestingly, just a few days earlier, Jin Yuzhi, CEO of Huawei's Intelligent Automotive Solutions BU, stated in an interview with the media that for future L3 and L4 autonomous driving, vehicles must have LiDAR. It seems that this debate over autonomous driving technology routes will continue for a long time and is far from over.

Trouble Stirred by the Pure Vision Route?

Probably no one anticipated that an apparently ordinary accident would trigger such widespread public discussion and the heightened attention of regulatory authorities.

On the evening of March 29, a Xiaomi SU7 was involved in a serious traffic accident while driving on the Deqing-Shangrao Expressway, Chiqi section. The incident quickly escalated, and on the evening of April 1, Xiaomi Automobile responded and released some information about the accident. As of now, the accident is still under investigation, and the official police report has not been made public.

Besides issues with the car door and battery, what has attracted the most attention is the vehicle's intelligent connected technology. Many believe that before the accident, although Xiaomi had been relatively conservative in promoting its intelligent assisted driving technology, some automakers had exaggerated or even misled the situation, which to some extent increased the probability of the accident occurring. In response, the regulatory authorities acted swiftly. On April 16, the Ministry of Industry and Information Technology issued a notice requiring automakers to thoroughly conduct "combined driving assistance" testing and validation, clarify system functional boundaries and safety response measures, refrain from exaggeration and false promotion, and strictly fulfill the obligation to inform. Specifically, it includes standardizing promotion, not expanding promotional norms and technical terms, and using the term "(combined) assisted driving" in accordance with automation classification standards.

The trend of returning intelligent assisted driving technology to rational promotion swept through every corner of the industry in a short period of time. During the 2025 Shanghai Auto Show, promotional slogans such as autonomous driving, intelligent driving, high-level intelligent driving, and infinite proximity to L3 autonomous driving almost disappeared, replaced by assisted driving and intelligent assisted driving, with repeated emphasis on "safety." Subsequently, media reports claimed that the promotional language of sales staff in many automotive brand 4S stores, including Xiaomi, had become more cautious overall, with some shifting to emphasize other selling points and only providing explanations on assisted driving when users actively asked about it.

In fact, this accident has also sparked industry discussions on specific technology routes. By sorting through and analyzing various information, it becomes clear that the standard version of the Xiaomi SU7 involved in the accident did not have LiDAR, and the pure vision mode has relatively poor recognition effect at night, which may result in insufficiently timely and accurate obstacle recognition. Additionally, when the vehicle's driving speed exceeds 100km/h, due to the relatively limited detection distance, there is a possibility that the AEB may not trigger. In this regard, some industry insiders believe that construction roads are generally marked on high-precision maps, and guidance for NOA rerouting will remind drivers to take over. If Xiaomi's intelligent assisted driving system gives a certain weight to real-time updated high-precision maps, there is a certain probability that the accident could have been avoided. However, it should be noted that road construction is divided into temporary and planned construction. Planned construction is generally marked uniformly on the road network, but temporary construction information may not be uploaded.

The Route Debate Reignites

In fact, the debate in the automotive industry about pure vision solutions versus multi-sensor fusion routes in the field of autonomous driving has never ceased. Automakers represented by Tesla have always promoted visual processing solutions, believing that the pure vision route is superior. However, some other automakers believe that visual processing solutions have many limitations that are difficult to overcome. In contrast, the fusion solution with LiDAR and millimeter-wave radar and other sensing hardware is an irreplaceable choice.

Before the release of Tesla's FSD V12 version, the pure vision route was always "the persistence of a few," and the multi-sensor fusion solution was favored by more automakers. However, as cost reduction pressures have gradually intensified in recent years, coupled with the rise and gradual maturation of technologies such as large models, driven by Tesla, the pure vision solution without LiDAR and high-precision maps has gradually become the mainstream.

The promotional boom of "no map" and "pure vision" peaked in August 2024: On August 26, HarmonyOS Intelligence held a new product launch event, officially launching the new AITO M7 Pro and AITO R7. Among them, the AITO M7 Pro is not equipped with LiDAR but instead carries Huawei's Kunlun ADS basic version system, adopting Huawei's self-developed pure vision technology; on August 27, XPeng Motors announced at its 10th-anniversary conference that the first model of the MONA series, M03, was officially launched, and also released multiple core technologies including the "XPeng Turing" chip and XPeng's new generation of "AI Eagle Eye Vision Solution"; on August 28, Geely Zeekr launched its intelligent assisted driving solution ASD, featuring "pure vision + end-to-end large model," priced at 4999 yuan.

It is worth emphasizing that the rise of "end-to-end" technology is not strongly tied to the pure vision route. The multi-sensor fusion solution can also be compatible with "end-to-end." However, due to considerations such as cost, coupled with the excellent performance demonstrated by Tesla's FSD, many automakers have enhanced their confidence in the "pure vision + end-to-end" technology route, believing that cameras alone can be used without the need for additional sensing devices such as LiDAR and millimeter-wave radar.

However, when FSD was at its peak, the industry did not entirely "lean towards" the pure vision route. "End-to-end is a software algorithm technology, and LiDAR is a hardware sensor. The two can coexist perfectly. Regardless of how software algorithm technology develops, for the realization of intelligent assisted driving, LiDAR is indispensable," a relevant expert from RoboSense bluntly stated in an interview with the media. The pure vision solution has certain limitations when the camera "cannot see" or "cannot see clearly," including but not limited to strong light exposure, low light environments at night, and the inability to distinguish between foreground objects and background colors that are the same, all of which can lead to accidents. Lin Jinwen, Vice President of Zeekr, also believes that "although LiDAR is indeed costly, the ultimate safety it can achieve in conjunction with algorithm software is something that pure vision cannot achieve. Not using LiDAR is simply finding excuses for cost reduction."

Fusion Perception Solutions are Safer and More Reliable

According to incomplete statistics, in 2018, there were nearly 100 global automotive LiDAR companies, but by last year, there were less than 10 left. However, on the path of continuous improvement in autonomous driving technology, LiDAR has never been completely abandoned and is even showing increasingly irreplaceable value and role amidst twists and turns in development.

Public information shows that the global passenger vehicle LiDAR market grew by 68% year-on-year in 2024, with the market size climbing to $692 million. Among them, Chinese LiDAR brands account for 92% of the market share. In terms of penetration rate, last year, the penetration rate of L2 and above assisted driving in Chinese passenger vehicles reached 55.7%. Zhang Yongwei, Vice Chairman and Secretary-General of the China Electric Vehicle Hundred People Council, predicts that this figure may approach 65% by 2025.

Huawei has always been a firm supporter of the fusion perception solution. In a recent interview with the media, Jin Yuzhi reiterated, "For future L3 and L4 autonomous driving, vehicles must have LiDAR." He explained that from a technical perspective, for any obstacle, the principle of LiDAR determines that it does not need to recognize the obstacle but can definitely detect it, allowing for braking or obstacle avoidance. In contrast, the pure vision route needs to recognize obstacles, which requires a lot of training. If there is not enough data, or even if the amount of data is very large, there will still be some unseen things. This kind of extreme and rare corner case where the obstacle is not recognized can bring risks. He also pointed out that cameras in pure vision solutions are limited in some scenarios, including very dark light and heavy rain. Vision essentially simulates human eyes, and unclear vision can bring risks. However, sensors such as LiDAR and millimeter-wave radar can play a role in extreme scenarios.

Pony.ai shares the same view. Pony.ai CEO Peng Jun also recently stated in an interview with the media that for L4-level autonomous driving technology, pure vision is difficult to achieve complete safety. He also pointed out another important reason, which is that after years of development, the cost of LiDAR has dropped significantly. "Earlier, everyone believed that LiDAR was too expensive to be used in cars, but the situation is now completely different." He pointed out that the current price of LiDAR is basically in the same range as cameras. According to the relevant expert from RoboSense mentioned above, the cost of currently mass-produced LiDAR has dropped to less than $500 (approximately RMB 3500), and it is expected that by 2025, the cost of some new products will fall to less than $200 (approximately RMB 1400).

It is worth mentioning that considering cost alone, it is unclear whether pure vision is more "cost-effective" than the fusion perception route. Because in the long run, although a portion of the hardware cost per vehicle is saved, automakers need to invest more research and development resources in the pure vision route. Shi Yezhou, the strategic head of Hesai Technology, said that comprehensive cost includes not only hardware costs but also the underlying research and development services and resource investments, including hidden costs such as algorithms, road testing, cloud computing, data annotation, simulation training, and system software.

All Roads Lead to "Rome"

There has never been only one path to "Rome." In fact, even within the fusion perception route, there is another option that does not require LiDAR – "full vision."

It is understood that "full vision" refers to removing LiDAR but retaining the positions of millimeter-wave radar and ultrasonic radar on the vehicle. Among them, millimeter-wave radar serves to detect the position and speed of distant targets and is not affected by adverse conditions such as light and weather, with strong anti-interference ability. Ultrasonic radar is used for close-range target detection and has high accuracy, providing accurate environmental information for vehicles in low-speed driving and parking scenarios. One for long-range and one for short-range, they serve as "assistants" to the core sensing component, the high-definition camera. Some experts point out that this is a "compromise" solution made by domestic automakers amidst the dilemma between cost and technology. For example, the visual solution of NIO's Lido L60 is equipped with a Centennics 4D millimeter-wave radar, with a maximum detection distance of 370 meters, which is stronger than ordinary millimeter-wave radar in terms of detection accuracy, active safety capabilities, and ability to cope with harsh weather conditions.

Meanwhile, Li Auto has also "blazed a new trail" under the existing technological conditions and levels. On May 7, Li Auto launched the second season of Li Auto AI Talk – "Li Auto VLA Driver Large Model." According to Li Xiang, CEO of Li Auto, VLA can see the physical world in its entirety through a combination of 3D and 2D vision. Coupled with VLA's complete brain system and capabilities in language and CoT (Chain of Thought) reasoning, it can not only see but also understand and truly execute actions, aligning with human operation methods. In the view of Li Auto, the VLA driver large model can solve the problem of fully autonomous driving, but it does not rule out the possibility of more efficient architectures emerging in the future. He believes that there is a high probability that more efficient new architectures will emerge in the future. After all, VLA is still based on the Transformer (deep learning model architecture), and it is still uncertain whether Transformer is the optimal structure.

As the "crown jewel" of smart cars, fully autonomous driving is by no means achievable in a short period of time, and this debate over the vision route versus the fusion perception route will continue. Perhaps, as Li Xiang said, other superior solutions may emerge in the future, and we cannot rule it out.

Note: This article originally appeared in the "Hot Spot Tracking" section of the June 2025 edition of "Auto Review" magazine. Keep an eye out for more.

Image: Sourced from the internet

Article: Auto Review

Typesetting: Auto Review