As NOA and L2+++ Expand, Autonomous Driving is Now a Race against Time

This article explores the company DeepRoute.ai and traces how the competitive benchmarks in autonomous driving are shifting. For a long time, autonomous driving was framed as a problem of seeing things accurately. But the harder challenge on real roads is how to understand what you see and translate it into action. The foundation model and Physical AI that DeepRoute.ai presented at Auto China 2026 stood precisely at this inflection point. DeepRoute.ai's choice - looking in the same direction as Tesla but delivering it in a form OEMs can actually use - poses a weighty question for Korean OEMs and Tier 1 suppliers as well. The autonomous-driving race is now moving from asking whether the technology is possible to asking when that possibility will reach mass production. And the time left to answer that question is less than one year.
 





When Maxwell Zhou stepped onto the floor for his keynote, his expression didn't look as relaxed as his casual attire suggested.
At the DeepRoute.ai booth keynote during Auto China 2026, he appeared a bit tired - perhaps from the overwhelming scale and packed schedule.
"I didn't think this would be such a long fight at first. Looking back, it really has been a long journey."

It has already been seven years since he completed his AI doctorate at the University of Texas at Dallas and co-founded DeepRoute.ai in Shenzhen in February 2019. His reflection on that trajectory was a monologue leading up to the next declaration - one poised to reshape the entire direction of the autonomous-driving industry.
"To improve autonomous-driving safety by 10x or 100x, it is impossible without large foundation models."






 
The history of autonomous-driving technology is, in truth, a history of 'separation.' Perception, Prediction, Decision, Planning - each function was implemented as an independent model, and engineers stitched them together with rules. When a camera detected a pedestrian, that information was passed to the prediction module, then the decision module made the call, and the planning module generated a path. It was an elaborately assembled pipeline. But within that sophistication lay a trap. Every time modules were joined, information was lost, and whenever a situation arose that the rules hadn't anticipated, the system made mistakes.
DeepRoute.ai recognized this structural limitation from its founding. That is why it was the first to adopt an Early Fusion Network. This approach - merging information at the raw sensor-data level - was aligned with the early currents of BEV (Bird's Eye View) and Transformer-based autonomous-driving architectures, and was a rare choice in the industry at the time. In 2022, DeepRoute.ai introduced a two-stage neural-network architecture that realized navigation using only SD maps, and in 2023 it launched China's first Map-free solution.
Then, in 2024, it succeeded in deploying an End-to-End (E2E) model on mass-produced vehicles. Camera footage goes in, driving actions come out - a so-called 'Vision-in, Action-out' architecture in which the intermediate module pipeline vanishes. By 2025, it had reached mass production of a VLA (Vision-Language-Action) model that connects visual input and decision-making within a single model. It was the first among China's independent autonomous-driving solution providers (excluding OEMs) to achieve E2E mass production, and it became the only company to reach VLA mass production.
In truth, this trajectory is not unique to DeepRoute.ai. Tesla has already walked the same path. The difference is that DeepRoute.ai is a third-party supplier that provides solutions to OEMs. Where Tesla binds data, models, and vehicles into a single closed loop internally, DeepRoute.ai ports that capability into a form that external OEMs can use with their own technology and infrastructure. The data stays within each OEM; DeepRoute.ai's model runs on top of it.
"This is not a model that simply performs specific functions - it is a model that understands and interprets the real world. Very few companies besides Tesla have actually implemented such a complete End-to-End architecture."

Looking in the same direction, but attempting a different implementation. That is the position DeepRoute.ai defines for itself.


  It would be a mistake, however, to read DeepRoute.ai's story as a smooth success narrative. What CEO Zhou shared on the Auto China stage was not just a list of achievements.
"In 2025, we tried to apply large models to vehicles, but we realized it was impossible with the existing approach. And we concluded that we needed an entirely new organization and entirely new talent."

In 2025 - at a time when DeepRoute.ai had already achieved E2E mass production and secured a leading market share in China's third-party ADAS market - the company made the radical decision to move past its conventional development methodology toward a new AI-driven paradigm. The judgment was that a traditionally siloed functional organization could not handle the complexity of foundation models. The conclusion that 'the organization itself must be transformed into an AI-native one' led to a wave of new talent recruitment.
The signal flare was the arrival of Chong Ruan, former Chief Scientist of DeepSeek. DeepSeek was the Chinese AI research lab that shook the global AI industry in early 2025 by achieving GPT-4-class performance with a small team and limited resources. The fact that he moved to an autonomous-driving company carried messages on multiple levels. First, that DeepRoute.ai is redefining autonomous driving as a pure AI research problem. Second, that an efficiency-centric model-development philosophy - squeezing high performance from limited resources - is being transplanted into autonomous driving. That is precisely what DeepSeek demonstrated in the AI industry.
Changes in the technological environment also supported this decision.
CEO Zhou said, "Until 2025, multimodal models were not sufficiently mature. But entering 2026, major technological breakthroughs occurred at Gemini, OpenAI, and DeepSeek."

DeepRoute.ai did not wait for the technology to fully mature - it changed its organization and development methods first. And the advances in multimodal models in 2026 became the driving force that elevated that decision into an actual mass-production strategy.


 
The key issue is that DeepRoute.ai itself was the first to recognize that E2E models are not the complete answer. Drive hundreds of kilometers and critical scenarios that a small model cannot comprehend will inevitably arise. A bus suddenly stops at an empty intersection and a pedestrian darts out from behind it. A blind spot at a Y-junction. A sign marking time-restricted bus-only lanes. It is structurally difficult for a small model operating at the level of pattern recognition to understand such situations in context.
"It works well in specific situations, but it is not safe in every situation, every city, at every hour," CEO Zhou said.

Currently, their ADAS scored 95.2% in the C-NCAP 2024 evaluation, earning five stars. The starting point is high. The question is what comes next.
A look at the Chinese market makes this reality even clearer. The majority of players still rely on small models because development costs are low and production speed is fast. In the short term, it looks like a rational choice, but DeepRoute.ai's assessment was different: the safety ceiling achievable with small models is already visible, and the only way to break through is with a foundation model.
The foundation model DeepRoute.ai has adopted takes a multimodal model with billions of parameters as the bedrock of autonomous driving. This model goes through multi-layered processes: pre-training, post-training, and driving-data training. It is not merely performing a specific function - it is a model that understands and interprets the physical world.
Chong Ruan organized this architecture around three capabilities: Driving, which controls the actual vehicle; Analyzer, which describes the vehicle's behavior in language; and Critic, which judges whether behavior is good or bad. Crucially, these three are not separate models. A single foundation model possesses all three capabilities simultaneously.
If the overall foundation model is on the order of 40 billion parameters, the Driver Model actually deployed in vehicles is compressed to 7 billion parameters or fewer - extracting and compressing only the necessary functions.
The Analyzer does not operate during driving. It plays a core role at the training and debugging stages. Its ability to analyze and explain why the vehicle made a particular judgment in a given situation determines the speed of model improvement.
Chong Ruan said, "The real competitive edge is not the model itself but how quickly you can iterate and improve."

The data-collection - training - testing - correction cycle, which used to take more than 100 hours, has been shortened to around 10 hours through foundation-model-powered pre-analysis and cloud-based simulation.





Another axis DeepRoute.ai emphasizes is 'Physical AI' - the ability to go beyond simple image recognition and understand and generate the physical dynamics of the real world. CEO Zhou explains it this way:
"If you look at the recent video-generation AIs, they follow real physical laws to an incomparably greater degree than a year ago. If you can generate it, you understand those physical laws."

This philosophy is directly reflected in their VLA architecture. VLA processes BEV imagery and LiDAR data through a visual encoder, combines them with a text encoder, and integrates everything into a single model. The output branches into two: a Trajectory Decoder that generates trajectories, and a Text Decoder that outputs Chain of Thought (CoT) reasoning and actions in language.
It is this Chain of Thought capability that enables Defensive Driving. When a vehicle in an adjacent lane decelerates suddenly without signaling, the VLA does not simply hit the brakes. Instead, it reasons through a chain of thought: 'Unsignaled sudden deceleration - temporary stop of bus ahead on the right - high probability of pedestrian emergence - decelerate to prevent collision.' It is the model structuring in language and executing what a human driver does unconsciously - slowing down based on an uneasy hunch.
Scenario Semantic Understanding also operates on this foundation. Time-restricted bus-only lanes, auxiliary lanes, speed bumps, construction zones - road conditions composed of text symbols are understood in context and used to determine routes.
Voice Control also derives naturally from this architecture. When a driver says things like "closer to the car ahead" or "a little faster," it adjusts longitudinal distance or acceleration; if the driver requests "prevent lane cuts," it changes the lane-change strategy itself.
A natural follow-up question at this point is how much of the VLA and the foundation model is DeepRoute.ai's own development. There was also a question about whether it is based on open source.
"We are not based on open source. We built it from scratch. Open source has many constraints and cannot change the mainstream. That's why we did it this way from the start. But nobody believes it. They just don't believe it,"
CEO Zhou added with a laugh.






 
Talk is talk, and a car is a car. There is only one way to verify how a presentation works on real roads. The live vehicle test began in the parking lot outside Hall A2. An 'NOA' indicator appeared on the right side of the instrument panel - a signal that NOA had been activated.
At the same time, logs began scrolling up on the left screen. Someone asked:
"Is this being stored in the cloud?" A DeepRoute representative riding along answered, "No. This is a monitoring tool. It is not a commercial feature." The data stays in the car.

The car drove for one hour. Beijing's downtown roads checked off the list of critical scenarios that DeepRoute had described in its presentation, one by one. The car picked its lane in a three-lane rotary. At interchanges with frequent merging traffic, it coordinated speed and spacing. 
Intersections where pedestrian signals and vehicle signals were not properly followed - a not-uncommon sight in China - continued to appear. When necessary, the car slowed or stopped and waited.
The road back to the exhibition center was gridlocked. Even in congested sections, the car stopped and flowed on its own. There was no intervention, and it felt more comfortable than when I drive myself.
It is difficult to conclude from a single test drive exactly what level of contextual understanding the foundation model delivers, but at a minimum, this car matched the slides in the presentation hall.






 
Chong Ruan did not talk only about technology at Auto China. Toward the end of his presentation, he raised the issue of organization and people. True to his DeepSeek background, his gaze shifted from model architecture to the structure of human-AI collaboration.
"In the past, we used individual small models for each function - pedestrian recognition, traffic-light recognition, and so on. As the number of models grows, the burden in terms of development, management, and headcount becomes enormous. Now people need to understand and utilize new technologies. But you must not use tools as your 'brain.' Tools are, at the end of the day, just tools."

This is more than organizational theory. To implement foundation-model-centric autonomous driving, engineers must take on the role of verifying and steering AI's decisions. In a world where AI becomes the driver, humans become the coach who trains that driver. This is why DeepRoute.ai says it needs 'superhero-caliber talent' - people who can handle models with tens of billions of parameters, design data pipelines, and build feedback loops. Such talent is exceedingly rare worldwide. Chong Ruan himself is that signal.
The fact that a researcher who built top-tier models with limited resources at DeepSeek has moved to an autonomous-driving company demonstrates that this industry is no longer an automotive-engineering problem - it has become the front line of AI research. And DeepRoute.ai is moving intentionally in that direction.


  No matter how right the technological direction may be, it is meaningless if you cannot survive in the market. There is a number DeepRoute.ai obsesses over alongside the technology debate: the cumulative count of mass-produced vehicles equipped with its system.
As of the end of 2025: a cumulative 200,000-plus units, over 35,000 units per month. 
As of October 2025, roughly 40% market share among China's third-party ADAS suppliers, with a year-on-year growth rate of 5X. Mass-production partnerships are underway with five OEMs - GWM, Geely, Leapmotor, and others - and more than 40 vehicle models are in the pipeline. 
Sales of GWM's Wey Lanshan model surged more than 3X after DeepRoute.ai's system was integrated.
The targets: a cumulative one million units by end of 2026, two million by end of 2027.
Behind the numbers lies a structural logic. In autonomous driving, data feeds the model, the model raises performance, and that performance in turn attracts more mass-production contracts. 
A single remark during the live test drive compressed this:
"Once a million units are deployed, you have data from a million units. Two million units, two million units of data. That is the difference."

So can this equation be transplanted to the Korean market? 
The question came up directly during the ride-along. For instance, what structure would a collaboration with Hyundai Motor take?
"We can take the model. But we cannot touch the data."

What DeepRoute.ai takes from Chinese OEMs is the model and development experience, not the raw driving data itself. Data generated in Korea must remain under the control of Korean OEMs, and sensitive information can only be shared in a limited fashion after filtering. It was here that the comment, "Ultimately the risk is whether data is exported or leaked," surfaced.
The feasible solution for now is a collaboration structure that is essentially a black box. The OEM receives the data first and opens it only within the necessary scope. The interior is never fully exposed.
Even within these constraints, the Map-free strategy functions as a realistic solution for global expansion. Because it does not rely on HD maps, there are no map-production costs, and it can be deployed immediately in regions where map updates lag. The opening of a Germany office and expansion into the GCC region are extensions of this strategy. And in this context, Korea becomes a natural point on the same coordinate.





 
The separate interview took place in the second-floor meeting room of the DeepRoute.ai booth. It was a glass-walled reception room. The exhibition floor across the way was visible at a glance.
Directly in the line of sight sat the Xiaomi booth - the company that, just three years after declaring it would build cars, mass-produced the SU7 and shook the market. 
DeepRoute.ai and Xiaomi faced each other across a single pane of glass. Xiaomi built cars and layered software on top; DeepRoute.ai builds only software and places it on other companies' cars. Two companies walking toward the same destination of autonomous driving from completely opposite directions.
To understand DeepRoute.ai's positioning, two companies provide useful comparison points: Waymo and Xpeng. Waymo has long been regarded as the benchmark for autonomous driving - Level 4 technology, HD-map-based, operating a small fleet of high-cost vehicles in select cities. But this model never translated into mass production. The Cruise case, in which financial support was cut, illustrates the limits of what happens when technology, however advanced, fails to connect to commercial scale.
It is against this backdrop that DeepRoute.ai placed mass production at the center of its strategy from the start. Data comes from scale, and scale comes from mass production.
Xpeng is another reference point. As CEO Zhou put it, "Xpeng is making a certain level of progress." Xpeng is one of the few companies that is both an OEM and has internalized its own autonomous-driving technology.
However, for an OEM to build a foundation model in-house, it must cross the critical threshold of talent and compute resources. Globally, only a handful of OEMs can clear that threshold. For most OEMs, a third-party supplier like DeepRoute.ai remains the realistic option.
The competitive landscape is divided into low-, mid-, and high-end segments. As one person on-site put it, "If Momenta eats up to mid, it can't go high-end. Then that's our territory." DeepRoute.ai's focus on foundation models and VLA is because technological superiority defends the premium segment.
Time pressure is also at play: Level 2++, and +++ must be completed by 2027. For most OEMs, the remaining time is not long. In reality, they have less than a year to set a direction and produce results. What is being tested now is not simply technology. It is whether a foundation model can achieve 10X the safety level of a small model in a mass-production environment, whether the flywheel actually works within a structure where data remains inside the OEM, and whether this equation holds globally beyond China.
Korean OEMs, too, stand before this question. 
Choosing not to set a direction is itself a choice - but that choice, too, carries a cost. DeepRoute.ai's bet is large, and its direction is clear.





 
Where is DeepRoute.ai headed? 
Zhou has laid out DeepRoute.ai  s long-term vision:  "We aspire for the company to become the AI infrastructure of the physical world in the future. Just like telecommunications and power utilities, we will grow into a foundational capability that underpins the operation of the real world. When it comes to intelligence in the physical world, DeepRoute.ai will serve as an essential part of this foundational capability system."


 

AEM(오토모티브일렉트로닉스매거진)

---