Is VLA a rival to Tesla V13?

At the end of February, Tesla's FSD entered the Chinese market, making all domestic intelligent driving companies feel extremely nervous.

In March, both "failures" and "smooth performances" of Tesla's FSD emerged. Although the smiles returned to the faces of the industry leaders, their inner anxiety remained unknown.

Looking back at the performance of Tesla's FSD after entering the Chinese market, there is a sense of dissonance, like putting "A Showcase of Baffling Intelligent Driving Behaviors" and "A Demonstration of an Excellent Intelligent Driving Student" side by side. On one hand, in Shanghai's Lujiazui, FSD used the bus lane as an overtaking lane and sped like crazy. Under the Canton Tower, when the navigation clearly showed a green light, the car suddenly braked sharply because it mistook the "Road Construction Ahead" warning sign for an advertising billboard. On the other hand, in normal driving, it demonstrated the skills of an experienced driver, accurately avoiding non - motor vehicles and smoothly completing maneuvers such as U - turns and turns, providing users with a proficient and smooth experience.

FSD performs smoothly in normal driving

The reason for this situation is quite simple. Tesla has a solid foundation in underlying AI technology, which can support the smooth experience of intelligent driving in normal use. However, since it is equipped with an end - to - end model (compared with the traditional modular intelligent driving model, although the end - to - end model does not cause information loss in the processes of perception, decision - making, and control, in essence, it still performs corresponding operations according to instructions and cannot understand the driving rules of special road conditions such as tidal lanes) and lacks a Chinese data training system, FSD does not understand the complex scenarios of the game between people and vehicles in China and cannot understand the driving rules of special roads. This has led to Tesla's "Showcase of Baffling Intelligent Driving Behaviors".

The root cause of the collective anxiety of the industry leaders lies here. Due to its powerful underlying logical ability, once Tesla makes up for its domestic shortcomings, it will surely cast a "shadow" over domestic car manufacturers.

At this time, VLA emerged.

One Good News and One Bad News

The good news is that the VLA model can address the "shortcomings" of the end - to - end approach. It integrates three actions: seeing, thinking, and doing. It uses cameras and lidar to collect road condition information, which is like equipping the car with a pair of 24k big eyes. The large language model analyzes the upcoming road conditions by solving problems such as traffic signs and pedestrian intentions. It's no exaggeration to say that it can even detect when a pedestrian is about to run a red light and can handle tidal lanes and bus lanes. Based on what it sees and thinks, VLA can plan the optimal route for the vehicle, control the vehicle, and even kindly explain the decision - making logic, such as "slowing down because a child suddenly rushed out ahead"...

To put it simply, VLA integrates vision, language, and action, endowing the vehicle with a "human - like thinking chain", evolving from the end - to - end "telling a story based on pictures" to "reading comprehension".

Since VLA can solve the technical shortcomings of the end - to - end approach, why are domestic car manufacturers still anxious about Tesla's FSD entering China? It's actually easy to understand. The basic and learning abilities of Tesla's FSD are very strong. By simply using video clips of Chinese roads found on the Internet for training, the system can already show the skills of an experienced driver. Once Tesla makes up for its data gap in China and overcomes its "acclimatization" here, it may become one of the best intelligent driving systems in the Chinese market. At a recent conference of hundreds of people, Zhou Guang, the CEO of DeepRoute.ai, and Wu Yongqiao, the president of Bosch China, also said bluntly that FSD V13 is one generation ahead of domestic high - level intelligent driving systems in end - to - end intelligent driving technology.

So, can domestic VLA bridge the gap in underlying technology? Both FSD V13 and VLA are regarded as products of the intelligent driving large - model stage in the industry. In Zhou Guang's view, the capabilities of VLA make it a "generalist system", that is, it has wide - ranging scenario adaptability in the vertical field.

"VLA can make up for the shortcomings of the end - to - end model and is a generalist driver. It can understand semantic information and the driving rules of special lanes. Only by becoming a generalist driver first can one become an expert in the driving field, that is, achieve full - scale autonomous driving." Zhou Guang regards the VLA architecture as a turning point towards L5. He believes that the essence of VLA is to build a spatio - temporal unified cognitive framework, which provides underlying support for achieving L5 - level autonomous driving.

In short, Zhou Guang believes that the VLA promoted by DeepRoute.ai is an optimal solution for moving towards L5 while retaining core AI capabilities, which meets both technological pursuits and commercialization needs.

The good news is exciting enough, but the bad news makes people a little uneasy - currently, there are no fully VLA - equipped models on the market. But don't worry, it's already in the works.

The "Comeback" Paths of Four Routes

Currently, there are four players in the domestic market that have clearly laid out VLA: Li Auto, a data "maniac"; Chery, a major player in joint - ventures; Geely, a "buying spree expert"; and DeepRoute.ai, a radical player.

Li Auto uses "MindVLA", which integrates high - end technologies such as 3D Gaussian coding and MoE Mixture of Experts architecture. Its decision - making accuracy in complex road conditions does lead its peers. However, the key problem is that they have to maintain both end - to - end and VLM systems simultaneously. Although they have sufficient data (confidence) - the coverage density of their dynamic data lake ranks first in the industry, and they boldly claim to achieve "data freedom" in 2025 - the R & D cost has skyrocketed, almost reaching the cost of a new car - making force.

Chery's trick is having many friends. It joined hands with Huawei and NVIDIA to develop the Falcon Intelligent Driving system and plans to implement the VLA model on the Falcon 900. However, the prediction ability of its World Model (WM) has not met the standard, and the launch time is set for 2027.

Geely, the "buying expert", uses the "Thousand - Mile Vast" intelligent driving system as its spearhead and launches a combined punch of "Full - scale AI + Heaven - Earth Integration", trying to overwhelm single technologies with an ecosystem.

DeepRoute.ai, the radical player, is fundamentally different from the previous three. It is not only one of the very few intelligent driving solution providers investing in VLA R & D but also has left the others behind and entered the mass - production stage. Without any surprises, models equipped with DeepRoute's VLA will be on the road by mid - year.

While others are just starting to research, DeepRoute is already in mass production? Actually, by taking a closer look at DeepRoute.ai's development path, we can find that it has always been at the forefront silently. In 2020, DeepRoute.ai proposed "map - less" intelligent driving, and it wasn't until 2024 that the trend of map - less driving emerged. In 2024, when car manufacturers were talking a lot about end - to - end technology, DeepRoute.ai had already put its end - to - end solution into mass production. In 2025, when everyone is competing for the end - to - end market share, DeepRoute's next - generation VLA is already in the process of mass production.

Zhou Guang once said bluntly that he is the "enlightening teacher" of domestic intelligent driving technology. Looking at it now, it's really not surprising.

In any industry, it is a common understanding that falling behind means getting beaten, and the intelligent driving industry is no exception. Car manufacturers all want to be the "most" advanced. At this time, the advantage of choosing a "mature and reliable" third - party becomes obvious.

Companies that have achieved mass production have built a technological moat first. Relying on the long - term reasoning and global decision - making capabilities supported by VLA technology, the actual usage frequency of urban NOA will be significantly increased, which in turn will enable the company to accumulate a larger - scale and higher - quality real - world scenario data. Based on the Scaling Law, the increase in data scale will feed back into the algorithm performance iteration, further improving the user experience.

For example, DeepRoute.ai has not only reached the mass - production stage but also established a cooperation with Qualcomm. Through in - depth research at the operator development level, there is a chance in the future to deploy the VLA model on more chip platforms, supporting both pure - vision and lidar versions. This means that as a third - party, DeepRoute has vividly demonstrated what it means to be "skilled, low - maintenance, and compatible with everyone".

Of course, it's not impossible for latecomers to catch up. But this means not only spending a large amount of money and resources. The most worrying thing is that by the time they finally succeed after great efforts, others may have already iterated to the third or even fourth generation, missing the best "golden window period" for development.

In conclusion, it's a thankless task. Just as the writer Zhang Dai said, "Don't raise a cow just to have a sip of milk." Since there are already mature pastures with mature processes from breeding (underlying logic) to care (extension ability of AI technology) to production (mass - production ability), there's really no need to do it on your own.

When Tesla announced in 2023 that FSD Beta V12 (Full - scale Autonomous Driving Test Version) would fully switch to the end - to - end architecture, it subverted the industry's perception to a certain extent. When Tesla synchronized the non - fully - fledged V13 to China, the domestic intelligent driving industry was stirred up again. The executives of car manufacturers are constantly vigilant about whether their hearts can withstand the impact on their own products after Tesla makes up for its data gap.

Currently, the leading domestic companies in the intelligent driving field have all achieved mass production, and their engineering capabilities are at the same level.

The competition in intelligent driving has shifted from engineering capabilities to the underlying capabilities of AI models. It can be expected that when domestic models equipped with VLA technology gradually hit the road this year, there may be a new answer to the question of which is stronger, Tesla's FSD or domestic VLA.