Running red lights, lane change violations... Tesla FSD struggles in China, where does the problem lie?

Tesla's FSD (Full Self-Driving) feature has finally taken a key step in the Chinese market.

On February 25, Tesla announced the launch of the city road Autopilot feature for users who ordered the FSD smart driving assistance function, with Tesla updating the software for Chinese customers in batches. This move marks the official entry of the highly anticipated FSD feature into the Chinese market.

However, the FSD feature being pushed out this time is not the fully autonomous driving capability that Musk claims can achieve ten times the safety of human drivers, but rather an optimized version based on EAP (Enhanced AutoPilot). According to the latest definition on Tesla's Chinese official website, the "FSD automatic driving assistance package" has quietly replaced the previous term "fully autonomous driving."

The rushed rollout of FSD, which has not yet evolved into its complete form, may indicate that Musk is indeed anxious in the face of Tesla's sales growth bottleneck and fierce competition from domestic automakers. Initial user feedback has exposed significant shortcomings in the localization of the FSD China version, including poor adaptation to traffic regulations, such as changing lanes while crossing solid lines and running red lights, indicating that there is still a long way to go before achieving true fully autonomous driving.

Reflected in the capital market, investors are clearly not convinced, as Tesla's stock price plummeted more than 8% overnight, with its market value falling below $1 trillion.

First Wave of Feedback: Driving Capability is Fine, Localization is Essentially Zero

The version of Tesla's FSD in China includes features such as traffic signal recognition, navigation assistance, and intelligent lane changing, while the cameras can also recognize the driver's attention level. The above features are all advanced functions of FSD.

As soon as the FSD China version was launched, it was put to the test by numerous Tesla owners. After browsing online, I found several honest reviews.

One WeChat user stated, "The driving capability of FSD is fine, inheriting the smoothness of the U.S. version end-to-end, but there are many issues with adaptation, such as running red lights and changing lanes while crossing solid lines... 'Localization is essentially zero.'"

After challenging a difficult intersection in Beijing with FSD, "Phoenix Car Research Institute" gave a relatively positive evaluation:

The Chinese version of FSD has significant differences from the U.S. version, but its completion level is impressive, making it hard to imagine that such a level could be achieved solely through online video learning

The driving performance of FSD is confident and decisive, whether changing lanes or turning, it appears crisp and neat, while fully respecting road rights, strictly adhering to traffic rules, and showing courtesy to vulnerable road users such as pedestrians and cyclists.

FSD does not require drivers to keep their hands on the steering wheel but requires the driver's gaze to always remain forward. In terms of takeover strategy, if FSD takes the wrong route, it will not prompt the driver to take over but will attempt to replan the route by itself.

"Phoenix Car Research Institute" also mentioned localization issues, stating that FSD "understands driving well but doesn't quite understand China." It can handle turns and navigate complex intersections steadily, but appears somewhat confused in specific situations such as bike lanes and bus lanes. FSD still has unresolved issues such as three-point turns and autonomous gear shifting.

However, it is understandable that this intelligent driving system from the United States is just landing in China, and making mistakes upon arrival is normal.

After all, according to relevant Chinese regulations, Tesla cannot transmit the collected driving data from China overseas to train models, and the U.S. side restricts FSD training in China. As background information, since Tesla FSD switched to an "end-to-end" architecture in 2021, the system directly outputs control commands after inputting sensor data, relying heavily on massive driving data for training.

To solve the data collection problem, Musk revealed that Tesla can only train FSD in a simulated environment using videos of Chinese roads available on the internet.

According to automotive blogger @不是郑小康， this is akin to American drivers learning to drive in China by watching videos.

The Bottleneck of Fully Functional FSD: The Triple Constraints of Data, Computing Power, and Compliance

Some netizens expressed after seeing evaluations from various automotive bloggers: FSD still has some merit. With the current iteration speed, domestic FSD is expected to catch up to the U.S. version level within six months.

However, without the support of local Chinese data, how far Tesla FSD can go in the fierce intelligent driving competition in China remains uncertain.

Although Tesla collects data through online videos, the quality of this data is significantly different from that collected in the field. China's traffic regulations and road environments differ significantly from those in the U.S., such as bus-only lanes, traffic islands, waiting areas, and special road users like delivery personnel and food delivery riders, all of which pose higher demands on FSD Elon Musk has also mentioned that one of the biggest challenges for FSD in China is the recognition and handling of bus-only lanes. These complex traffic scenarios require a large amount of data for training support, but Tesla can currently only collect some data in limited areas through engineering vehicles in China.

In addition to the above issues, there is also a computing power bottleneck. Tesla has established the Cortex supercomputing cluster overseas for FSD training, equipped with about 100,000 NVIDIA H100 and H200 chips, but this computing power cannot be used for domestic data training.

According to media reports, due to policy impacts, Tesla is considering leasing or building a computing power center in China to meet the huge data training needs. However, the specific implementation and effectiveness of this plan remain unclear.

Musk is really anxious to bring the "stripped-down" FSD to China

The launch of the "stripped-down" FSD system in the Chinese market reflects Tesla's urgent need to address the sales growth bottleneck and fierce competition from domestic automakers.

Tesla's sales have hit a growth bottleneck. In 2024, Tesla delivered a total of 1.789 million new vehicles, a year-on-year decrease of 1.1%, marking the first decline in new vehicle sales for Tesla in ten years. In the critical Chinese market, the Model 3/Y, the two main volume models, have already shown signs of sales fatigue.

Faced with fierce competition from domestic automakers, Tesla urgently needs FSD features to enhance product competitiveness. However, do the issues that have emerged now indicate that the core capabilities of the Chinese version of FSD are severely mismatched with its high price of 64,000 yuan? In comparison, many domestic automakers offer intelligent driving features for free with vehicle purchases.

Tesla's launch of the "stripped-down" FSD in the Chinese market is undoubtedly a risk. Although the introduction of FSD in China is highly topical, hastily launching an immature version may affect Tesla's reputation. In the face of fierce competition from domestic automakers and a complex market environment, Tesla needs to resolve the data bottleneck to truly establish itself in the Chinese autonomous driving market