EN / 中文

End-to-End Autonomous Driving: Tesla, Huawei, XPeng and Other Players’ Divergent Technical Routes

by zhijiazuiqianyan·November 18, 2026

In 2024, end-to-end technology became one of the most closely watched directions in the intelligent driving industry.

From Tesla's FSD to domestic automakers, companies have successively launched end-to-end solutions. The industry once viewed it as a critical path to breaking through the bottlenecks of traditional intelligent driving architectures.

Two years later, end-to-end is no longer just a technical label for a few companies; an increasing number of mass-produced models are adopting data-driven driving models.

Although all are pursuing end-to-end approaches, different companies have not followed the same path forward.

Some are pushing VLA (Vision-Language-Action) toward mass production, others are introducing world models into their training systems, some are attempting to further optimize driving strategies via reinforcement learning, and still others are beginning to combine embodied foundation models with driver assistance.

After two years of development, where exactly has each company's end-to-end approach reached? What substantive changes have occurred in their technical solutions?

01. Tesla: Continuing to Deepen the End-to-End Model

If discussing the evolution of end-to-end technology, Tesla is impossible to ignore.

Tesla was among the first to apply large-scale end-to-end neural networks to FSD. Its core concept involves using real-world road data accumulated by its fleet to let the neural network directly learn the relationship between the driving environment and vehicle control behaviors, thereby reducing the numerous manually designed rules and intermediate processing steps found in traditional systems.

By 2026, Tesla continues to iterate along this path.

FSD V14, released this year, continues to upgrade around the end-to-end foundation model.

In its Q1 2026 update, Tesla disclosed that V14 further improved the reinforcement learning training phase to enhance handling of long-tail scenarios. It also upgraded the visual encoder and optimized model compilation and inference efficiency, with official claims stating runtime latency has been reduced by up to 20%.

As of now, FSD has entered the V14.3 iteration series. It should be noted that this discussion refers to FSD (Supervised), which still requires continuous driver supervision and does not imply that the vehicles have achieved fully autonomous driving.

A characteristic of Tesla's route is that it has not frequently changed technical names but has instead consistently iterated on data scale, neural networks, and training efficiency.

This indicates that end-to-end has not lost value due to the emergence of new technologies like VLA or world models. How to continuously improve existing models remains a very realistic technical direction.

02. Huawei: Extending from End-to-End to Cloud World Models and Online Reinforcement Learning

Huawei's changes are mainly reflected in the further upgrade of its training system and vehicle-side decision-making architecture.

In April 2026, Huawei officially released ADS 5 for its Qiankun Intelligent Driving system, utilizing the WEWA 2.0 architecture at its core. This architecture no longer emphasizes only how vehicle-side models complete perception and driving decisions but further strengthens the connection between cloud training and vehicle execution.

In the cloud, WEWA 2.0 introduces a multi-agent mechanism, allowing different traffic participants to form dynamic interactions within simulated environments. It also incorporates online reinforcement learning to attempt to improve model training and validation efficiency.

On the vehicle side, safety risk field technology is introduced, evaluating risks in different directions around the vehicle to provide a basis for driving decisions.

This is essentially an extension of traditional end-to-end systems.

In the past, models relied more on real-road data to learn driving behaviors. Now, Huawei is paying closer attention to the dynamic relationships between traffic participants and the potential risks associated with different driving actions.

For example, at complex intersections, vehicles must not only identify pedestrians and other vehicles ahead but also combine their motion states to judge potential interactions between different traffic participants.

Such problems are difficult to solve solely through static object recognition.

Therefore, Huawei combines world models, multi-agent mechanisms, and reinforcement learning, attempting to allow models to gain richer scenario experience during the training phase.

Looking at the technical changes in ADS 5, Huawei's focus is no longer just improving the driving capabilities of vehicle-side models but rather perfecting the entire system from cloud training to vehicle-side decision-making.

03. XPeng: Beginning to Enhance Temporal Understanding and Future Prediction

XPeng is one of the domestic automakers that began promoting end-to-end mass production relatively early.

In 2024, XPeng launched an end-to-end large model solution composed of the XNet perception network, XPlanner planning model, etc., gradually applying data-driven models to mass-produced driver assistance systems.

By 2025, XPeng further released VLA 2.0, advancing the driving model to a physical world large model with billions of parameters on the vehicle side, removing the language translation step, and capable of generating actions directly from visual signals.

In 2026, this solution has seen more substantive changes.

In April this year, XPeng published the X-World world model technical report, detailing the specific application of world models in end-to-end R&D.

X-World uses video generation technology to generate multi-view road scenes that may occur in the future based on historical footage from multiple cameras and vehicle actions. It is not only used for scene generation but has also entered R&D phases such as closed-loop simulation, online reinforcement learning, and data synthesis.

There is a crucial distinction here.

VLA 2.0 focuses on learning driving strategies from real driving data and directly generating actions; X-World focuses on deducing potential changes in the surrounding environment given specific actions, providing an interactive environment for training and validation.

The two are not mutually exclusive but serve different tasks.

By September this year, XPeng further upgraded the second-generation VLA through XOS 6.3.0, incorporating historical information, future prediction, and continuous decision-making capabilities into a new architecture.

According to public materials, XPeng's VLA 2.0 has entered mass production application, and X-World is already actually used in R&D and validation.

This means XPeng's technical focus has extended from simply training driving models to further exploring how to enable models to understand scene changes over longer time horizons.

04. Li Auto: MACH VLA Iteration and Chip-Model Synergy Become Key

Li Auto's end-to-end route has also undergone noticeable changes in recent years.

From the end-to-end model in 2024 to VLA in 2025, Li Auto has gradually combined vision, language, and driving behavior, attempting to give driver assistance systems stronger spatial understanding and behavioral decision-making capabilities.

Currently, Li Auto is focusing on advancing MACH VLA.

What is noteworthy about this solution is that Li Auto is not only expanding model capabilities but also simultaneously promoting the vehicle-side computing platform.

The self-developed M100 chip has entered mass-produced models, providing the hardware foundation for deploying larger-scale models.

Here lies a very practical engineering problem: while cloud models can continuously increase parameters and computational load, the intelligent driving system in vehicles must complete real-time inference within limited computing power, power consumption, and response time.

A larger model size does not necessarily mean better driving experience if deployed directly onto the vehicle.

Therefore, Li Auto needs to balance model architecture, inference efficiency, and chip adaptation simultaneously.

In delivery information announced in October2026, Li Auto confirmed that OTA 9.2 would begin rolling out in November, including the upgraded MACH VLA 2.5.

This indicates that Li Auto is continuing to push VLA mass production iterations rather than staying at the early technical release stage.

The key point of Li Auto's route is how to use synergy between models and self-developed chips to ensure more complex driving models run stably on mass-produced vehicles.

05. NIO: NWM World Model and Fully Closed-Loop Reinforcement Learning

NIO's technical route has a distinct characteristic: significant investment in world models and training loops.

After launching the NWM world model, NIO gradually integrated it with its existing driver assistance system, forming the NADArch 2.0 end-to-end architecture driven by NWM.

Public technical documentation shows that NWM possesses long-term temporal modeling capabilities, enabling better handling of complex road scenarios.

More notably, starting in 2025, NIO began switching its training system to a paradigm combining world models + closed-loop reinforcement learning. The mass-production version in January 2026 was the first to apply complete closed-loop reinforcement learning to driver assistance.

The modeled capabilities driven by NWM have also expanded from partial scenarios to comprehensive coverage of urban and highway navigation assistance.

By late January 2026, applicable models had begun receiving updates to a new NWM version.

Behind this technical change lies a question worth discussing.

Traditional imitation learning primarily relies on human driving data, where models form their own driving strategies by learning from drivers' operations.

However, human driving data cannot directly provide optimal decisions for all situations.

The same scenario may allow for multiple driving styles. Achieving a better balance between safety, efficiency, and comfort requires the model to learn and optimize further.

This is where the value of reinforcement learning lies. It can evaluate and optimize different driving strategies through environmental feedback, rather than simply imitating historical driving behavior.

NIO's current direction is to create a tighter training system involving world models, end-to-end architectures, and reinforcement learning. The focus of this route is no longer just teaching the model how to drive, but researching how to continuously optimize driving strategies.

06. Xiaomi: XLA Begins Introducing Embodied Foundation Models to Driver Assistance

Compared to the previously mentioned companies, Xiaomi's intelligent driving receives less public discussion, but its technical route changes remain noteworthy.

In March 2026, the new generation SU7 launched the Xiaomi XLA cognitive large model, marking Xiaomi's entry into a new architectural phase for driver assistance.

According to Xiaomi's official materials, XLA is built upon the self-developed Xiaomi MiMo-Embodied foundation model, emphasizing multimodal information acquisition, scene cognition, and behavioral reasoning capabilities.

Its technical approach differs from traditional end-to-end models.

In the past, driver assistance models were mainly trained on driving data, learning the relationship between the road environment and driving actions.

Xiaomi hopes to introduce broader spatial understanding and real-world cognitive capabilities into driver assistance through embodied foundation models.

It is important to note that an embodied model does not mean the vehicle possesses complete cognitive abilities similar to humans. Its actual value still needs to be verified through specific driving tasks and mass-production performance.

In terms of mass-production progress, the new generation SU7 is already equipped with XLA, while some models of the first-generation SU7 and YU7 require subsequent OTA upgrades.

What is worth observing about Xiaomi's route is its attempt to link automotive driver assistance with more general-purpose embodied intelligence models.

In the future, whether such models can bring actual capability improvements in complex road scenarios still requires verification with more mass-production data.

07. Momenta: R7 World Model Enters Mass Production Across Multiple Brands

Strictly speaking, Momenta is not an OEM but an intelligent driving supplier.

However, looking only at automakers' self-developed solutions might overlook significant progress in domestic end-to-end technology. Therefore, we include Momenta's technical solution here.

In April 2026, Momenta officially announced the mass-production debut of the R7 world model.

The technical system of this solution mainly includes world model pre-training, world model simulation, and reinforcement learning.

Its core idea is to enable the model not only to learn scenarios the vehicle has experienced in the past but also to predict potential future changes in the physical world.

For example, when driving at high speed, if an irregular obstacle suddenly appears ahead, the system needs to judge the obstacle's position, motion state, and changes in surrounding vehicles before deciding whether to decelerate or detour.

Such scenarios are difficult to cover in advance with traditional rules and are not easily collected in large quantities on real roads.

World models attempt to use existing data to learn environmental change patterns, generating and deducing more scenarios in virtual environments for model training and validation.

In practical application, R7 is no longer stuck at the technical demonstration stage.

Momenta's official website disclosed in November 2026 that R7 is entering more vehicle models, including related projects with brands such as Mercedes-Benz, BMW, Cadillac, and Audi.

This indicates that world models are beginning to extend from R&D systems to multi-brand mass-production projects.

Of course, entering mass-produced models does not mean capabilities have been fully validated in all scenarios, but it at least shows that world models are becoming an important technical investment direction for suppliers in the next phase.

08. After Two Years of End-to-End Development, Where Is the Real Gap?

Viewing these companies together reveals a clear change.

Two years ago, the focus of industry competition was whether end-to-end could be put into vehicles. Now, end-to-end is no longer unique to a few companies; various firms are adding new capabilities on top of it.

As for which route is best, there is no final answer yet.

It can only be said that they reflect several practical issues encountered by end-to-end technology as it develops today.

First, real driving data still has boundaries. Common scenarios can be accumulated in large volumes, but truly dangerous and rare scenarios are hard to obtain through natural vehicle operation. World models and simulation training are attempting to change how models acquire training scenarios.

Second, imitating human driving does not equal finding a better driving strategy. Models need not only to learn how drivers operated in the past but also to continuously optimize for safety, efficiency, and comfort. Reinforcement learning addresses precisely this issue.

Third, driving models require stronger scene understanding and future prediction capabilities. Roads are not static; dynamic interactions always exist between traffic participants. Simply improving object recognition cannot solve all complex driving problems. Temporal modeling, world models, and multimodal models are all attempting to supplement this capability.

Additionally, there is a non-negligible issue: the more complex the model, the harder it becomes to balance vehicle-side computing power, inference latency, system safety, and mass-production costs.

This is why, even though everyone is advancing end-to-end, the final technical combinations adopted differ.

Looking back, end-to-end has not become obsolete, nor has it been replaced by VLA, world models, or embodied models.

It is more like a critical step in intelligent driving moving from traditional rule-driven approaches to data-driven ones.

But reaching this point, the industry has gradually realized that merely teaching models how to drive is not enough.

What truly needs to be solved in the next phase is enabling models to make reasonable judgments when facing scenarios they haven't sufficiently seen; allowing models to not just imitate human driving but continuously optimize driving strategies; and ensuring these capabilities are truly implemented under limited vehicle-side computing power and strict safety requirements.

This is the deep reason why technical routes have begun to diverge after two years of end-to-end development.

In the future, assessing a company's intelligent driving level may require asking not just whether they have end-to-end, but what capabilities they have added on top of end-to-end, and whether those capabilities have truly translated into driving performance in mass-produced vehicles.