In the field of autonomous driving, end-to-end models have replaced traditional modular architectures over the past two years, becoming the mainstream technical foundation of the industry. However, end-to-end models primarily address the issue of information transmission efficiency; they do not inherently possess deep understanding capabilities of scenarios, nor can they proactively predict changes in future states.
Subsequently, the industry diverged into two technical enhancement paths: VLA (Vision-Language-Action) and world models. The former introduces language models as the reasoning hub, while the latter predicts the future by simulating environmental evolution internally. At GTC 2026, the divergence between these two paths was publicly brought to the forefront. This topic is also being discussed in the Autonomous Driving Frontier communication group; those interested are welcome to join the conversation.
However, entering the second half of 2026, practices from multiple leading enterprises indicate that VLA and world models are not mutually exclusive. One is responsible for understanding what to do now, and the other for predicting what will happen after taking action; the two can be fully integrated into a single framework. At the inaugural Embodied AI Foundation Model Deployment symposium at CVPR 2026, Tesla, Waymo, and XPeng shared the stage, and the core agenda has shifted to the construction of foundation models.
So, as VLA and world models move towards integration, will the next stop for autonomous driving models be foundation models?
What Is a Foundation Model?
Foundation models originated in the field of natural language processing, referring to large models pre-trained on massive datasets and adaptable to various downstream tasks. In simple terms, a large model first completes general education on large-scale data, followed by targeted fine-tuning for specific tasks, allowing the same model to adapt to multiple different applications.
End-to-end autonomous driving models have already achieved unified mapping from sensors to control. Foundation models take this a step further; they are no longer confined to the driving task itself but integrate capabilities such as language understanding and future prediction to build a general-purpose agent capable of deeply understanding the operating rules of the physical world.
Compared with traditional end-to-end models, the core difference lies in the fact that end-to-end autonomous driving models solve the mapping problem from pixels to control, whereas autonomous driving foundation models pursue a deep understanding of object motion laws, the behavioral logic of traffic participants, and scenario evolution trends.
As Cao Tongyi, CTO of DeepRoute.ai, said at GTC 2026, foundation models enable autonomous driving systems not only to execute driving actions but also to understand traffic scenarios, explain decision-making logic, and evaluate driving behaviors. From this perspective, autonomous driving foundation models do not replace end-to-end models but expand and upgrade upon them. End-to-end is an innovation at the architectural level, while foundation models represent a leap at the capability level.
Why Are VLA and World Models Moving Towards Integration?
From 2025 to early 2026, the most intense debate in the autonomous driving field was over which path was superior: VLA or world models. VLA (Vision-Language-Action) models introduce large language models as the reasoning hub within the end-to-end framework, enabling the system to understand scenarios and make interpretable decisions; whereas world models allow the system to simulate environmental evolution internally, predicting how the world will change after an action is taken.
The divergence between the two paths was once very sharp. Li Chuanhai, CTO of Geely Auto, and Cao Xudong, CEO of Momenta, among others, publicly questioned VLA. Jin Yuzhi, CEO of Huawei's Intelligent Automotive Solution BU, even called VLA a shortcut solution, splitting the autonomous driving field into two major camps: world models + reinforcement learning, and VLA large models.
In October 2025, Tesla released FSD V14, which integrated xAI's Grok large language model, achieving breakthroughs in navigation interaction and intent understanding, allowing users to complete route planning and other operations via natural language. Simultaneously, at the underlying technical level, Tesla utilized world model-related patents to generate training data and enhance human-like driving capabilities. This also demonstrates that in the autonomous driving field, VLA and world models are not an either-or choice; they can be developed simultaneously and together.
XPeng's statement at CVPR 2026 was also of landmark significance. Liu Xianming, head of its General Intelligence Center, explicitly responded that VLA and world models are not in a competitive relationship. Its second-generation VLA learns how to act from human driving behaviors, while the world model learns future states to predict how the world changes after an action. Together, they constitute the foundation model for the physical world.
Zhan Kun, head of foundation models at Li Auto, also expressed a similar view at GTC 2026: when vision, language, and action are unified into a single model, it is no longer just an autonomous driving model but is gradually evolving into a general-purpose agent oriented towards the physical world.
From the technological trends of these enterprises, it can be seen that VLA addresses semantic understanding and interpretable decision-making, while world models address future state prediction and deduction. The two solve problems at different levels and are naturally complementary.
The two paths are not substitutive; rather, they can jointly enhance the model's understanding and action capabilities regarding the physical world through different training signals, thereby building a sufficiently powerful foundation model for the physical world. Moving from the divergence of VLA and world models to integration, the autonomous driving industry is completing a critical leap from capability splicing to a unified model.
What Are the Leading Players Doing?
The competition for foundation models is now in full swing. Although technical paths vary among companies, the direction is highly consistent.
DeepRoute.ai released a 40-billion-parameter VLA foundation model at GTC 2026. The design concept of this model is a trinity, integrating three roles—driver, analyst, and referee—within the same architecture.
The driver is responsible for outputting driving control signals in real time based on visual input; the analyst is responsible for understanding the semantics of traffic scenarios, identifying key events, and providing causal reasoning; the referee continuously evaluates the safety, comfort, and naturalness of driving behaviors. The referee's evaluation results are also fed back to the model itself to optimize the next decision, allowing the system to self-reflect while executing.
Li Auto released its next-generation autonomous driving foundation model, MindVLA-o1, at NVIDIA GTC 2026. Based on a native multimodal MoE (Mixture of Experts) Transformer core architecture, the model constructs an autonomous driving foundation model oriented towards physical world intelligence through five major technological innovations: 3D spatial understanding, multimodal thinking, unified behavior generation, closed-loop reinforcement learning, and hardware-software co-design. Among these, the closed-loop reinforcement learning framework, combined with a high-fidelity simulator, enables the model to continuously explore and optimize strategies within the world simulator, reducing training costs by approximately 75%.
XPeng also fully disclosed its world model technology roadmap at CVPR 2026. Its physical world foundation model consists of two pillars: the second-generation VLA and the world model. The second-generation VLA has billions of parameters, is trained on over 100 million video clips, and each version of the model is trained on over 4 trillion tokens. In terms of the world model, XPeng released the X-Mind technical framework, which endows the system with visual chain-of-thought capabilities by embedding a predictive world model, allowing the vehicle to complete spatiotemporal deduction before executing actions.
Tesla showcased its foundation model path for robotics at CVPR 2026. The core concept is "three lines, one core," meaning autonomous driving, the humanoid robot Optimus, and digital agents share the same foundation model.
Supporting this path are Tesla's approximately 1.3 million vehicles globally equipped with supervised autonomous driving capabilities, with FSD accumulating over 10.8 billion miles driven, providing massive real-world driving data for foundation model training. In terms of model architecture, the inputs are multimodal information such as camera video, navigation instructions, and vehicle kinematics, while the output directly generates control actions, with no artificially defined module boundaries in between.
Waymo also unveiled its world model pre-training → mid-training → post-training framework for the first time at CVPR 2026. Instead of training from scratch, Waymo directly utilized Google DeepMind's Genie 3 general world model as a foundation, added Waymo-specific sensor data (multi-camera, LiDAR, millimeter-wave radar, HD maps) for mid-training, and finally performed fine-tuning and distillation for driving tasks. This also indicates that Waymo is transforming autonomous driving from a traditional software engineering problem into a foundation model problem.
From DeepRoute.ai to Li Auto and XPeng, and then to Tesla, although the paths vary, they all point in the same direction: foundation models are being transformed from a vague academic concept into engineerable and deployable technical solutions.
Although differences still exist, they are no longer debates over technical routes but choices of implementation paths. The respective technologies of these leading players all point to a common conclusion: foundation models, as a technical direction itself, are becoming the greatest common divisor of the industry.
What Challenges Do Foundation Models Face?
Although the direction of technological development is clear, foundation models still face multiple constraints from concept to large-scale deployment.
Computing power is an unavoidable topic. The parameter count of foundation models often reaches tens of billions or even hundreds of billions, and the demand for computing resources for training and inference far exceeds that of traditional solutions. However, the more pressing contradiction lies on the vehicle side. Electronic control systems require a decision response speed of 100Hz, while the inference speed of general VLA models is generally less than 10Hz, a difference of more than an order of magnitude.
To fully unleash the capabilities of large models under limited vehicle power consumption, it requires simultaneous advancements in the computing power of automotive-grade AI chips, model distillation and compression, and inference engine optimization. Cloud computing power and deployment frameworks solve how to train the model and how to deploy the system in the vehicle, while the issue of whether it can run on the vehicle side ultimately falls on chip and model efficiency. Regarding the lightweight deployment of models, Autonomous Driving Frontier previously discussed this in detail; those interested can check it out: Lightweight Model Deployment: How to Make Autonomous Driving Go from "Able to Run" to "Running Well"?
Data is another bottleneck. Training foundation models requires massive amounts of high-quality driving data to cover a sufficient number of scenarios, enabling the model to learn how to handle various long-tail scenarios. Tesla has over 10.8 billion miles of real-world driving data, but most enterprises do not possess such data scale, meaning they are naturally at a disadvantage in terms of model generalization capabilities.
Regarding the issue of data scarcity, synthetic data is highly anticipated. Autonomous Driving Frontier has also specifically discussed synthetic data: Can Synthetic Data Replace Real Data in the Training of Autonomous Driving Large Models? Of course, world models themselves can also be used to generate training data to compensate for the lack of real data. However, the distribution difference between synthetic and real data remains a problem to be solved; models trained on simulated data may struggle to adapt when deployed in real-world scenarios.
The interpretability and safety of models are even more unavoidable topics. The criticism of end-to-end models as "black boxes" will only be more pronounced in foundation models. The larger the model parameters, the more difficult it is to trace the decision-making logic. When publicly opposing the distillation approach, Huawei's Jin Yuzhi pointed out that hallucinations are fatal in driving scenarios. How to ensure the interpretability and safety of decisions while maintaining model capabilities is a threshold that foundation models must cross to reach mass production.
Final Thoughts
Returning to the initial question: is the next stop for autonomous driving models foundation models? Judging from industry trends in 2026, the route debate between VLA and world models is being replaced by integration, and the product of this integration is the foundation model. Tesla, Waymo, XPeng, Li Auto, and DeepRoute.ai—these players who once took different technical paths—are now converging in the same direction. As Liu Xianming from XPeng said at CVPR 2026: "Only companies capable of building foundation models have the potential to truly achieve L4." #AutonomousDriving #AutonomousDrivingModels #FoundationModels