From modular to end-to-end, the way autonomous driving understands the world has been completely revolutionized. Although modular architecture features clear logic, it struggles in complex urban traffic conditions. The frequent emergence of end-to-end approaches is precisely aimed at solving the problems inherent in modular architectures. But with the advent of end-to-end, has the modular architecture truly disappeared?
01. Is One-Stage End-to-End Still the Only Answer?
The evolution of Tesla's FSD has become a crucial factor driving the end-to-end route. Starting with FSD V12, Tesla significantly reduced traditional rule-based code, allowing neural networks to take on more driving decision-making tasks, which many consider a key milestone for one-stage end-to-end systems.
The so-called one-stage end-to-end approach starts from sensor inputs and directly generates driving behaviors through a large neural network, bypassing traditional independent modules such as perception, prediction, and planning. By 2026, the end-to-end technology itself is also evolving. Although Tesla's FSD V14 retains the basic end-to-end framework, it introduces Chain of Thought (CoT) capabilities into the decision-making process, generating a natural language reasoning chain before outputting the steering wheel angle.
The reason for this change at Tesla is that while pure end-to-end models perform well in complex scenarios, it is difficult for engineers to trace their decision-making rationale. Introducing Chain of Thought is equivalent to leaving a readable reasoning trace before outputting an action, which at least gives developers the opportunity to understand why the model makes a specific driving decision.
Along this direction, the industry has begun to systematically explore how to equip end-to-end models with deeper scene understanding capabilities. The concept of VLA (Vision-Language-Action) is to insert a language model between visual inputs and action outputs, leveraging the language model's grasp of semantics and logic to help the system understand the nature of the scene it is processing.
For instance, if a car is parked diagonally at an intersection, it might merely appear as an obstacle based solely on visual features. However, combined with language reasoning, the system can determine whether the car is waiting to turn left or has stalled and broken down, and the corresponding driving strategies for these two situations are entirely different. Nevertheless, the two conversions from vision to language and from language to action in VLA will bring additional computational overhead and response latency. How to strike a balance between reasoning depth and reaction speed remains a problem to be solved.
World models take a different approach. Instead of language reasoning, they enable the system to internally simulate and deduce what might happen in the next few seconds before executing an action. For example, if there is a bicycle swaying on the roadside ahead, the world model will deduce several possible motion trajectories and then select the safest response. This deductive capability is similar to the mechanism by which human drivers anticipate risks in their minds.
XPeng's X-Mind technical framework is an attempt in this direction, enhancing the system's predictive capabilities in dynamic environments by embedding a predictive world model. Both VLA and world models have made substantial progress in 2026, and the industry has gained a clearer understanding of the relationship between these two routes. XPeng explicitly stated at CVPR 2026 that VLA and world models are not in competition; rather, they jointly constitute the two pillars of the foundational model for the physical world.
The two address problems at different levels. VLA is responsible for semantic-level understanding of scenes, while the world model handles spatiotemporal-level deduction of dynamic processes. Combined, they enable the system to possess both scene judgment and risk anticipation capabilities.
From these industry attempts, it is evident that end-to-end approaches have taught autonomous driving how to drive. However, to cope with the complexity and uncertainty of the real world, the system still needs to establish deeper cognitive capabilities regarding the physical world. End-to-end solves the problem of driving skills, while VLA and world models address the issue of enabling the system to truly understand the world it is driving in.
02. Have Traditional Modules Truly Disappeared?
After so much background, we can now turn to today's main topic. Since end-to-end models can directly output driving behaviors, is there still a need for traditional modules such as perception, prediction, and planning? Has the modular architecture already disappeared?
To answer this question, we must first clarify one thing: what exactly has disappeared when we say the modular architecture is gone. If it refers to the software stack strictly divided by processes such as perception, prediction, planning, and control, then the modular architecture is indeed fading.
In 2026, the Chinese automotive industry is transitioning from modular architecture to end-to-end architecture. WeRide's WRD 3.0 uses a single large AI model to integrate perception, prediction, planning, and control; Bosch's one-stage end-to-end solution has also achieved mass production in China.
Mainstream intelligent driving solutions have shifted from the classic modular approach of perception, decision-making, and control to end-to-end neural networks. However, if we are referring to the engineering mindset of splitting an autonomous driving system into independently developable and verifiable components, it has not only not disappeared but has survived in a new form within the new architecture.
Let us first look at the actual changes that have occurred in the industry. At CVPR 2026, Wayve divided autonomous driving routes into three generations. AV1.0 is the classic modular stack, where perception, prediction, planning, and control each perform their respective duties; AV1.5 wraps a rule checker around the end-to-end model; AV2.0 tasks the end-to-end model with the primary intelligence from perception to control.
Wayve specifically emphasized one point: what truly matters is not whether there are rules, but that the responsible entity for safety assurance has changed. Traditional systems rely on rules written by engineers for safety, whereas the AV2.0 approach is to enable the model to learn risk assessment. However, even within the AV2.0 framework, the modular mindset still exists; the only difference is that the boundaries of the modules have shifted from external functional decomposition to internal component division within the model.
TIER IV's route illustrates the same point. In July 2025, the company released an end-to-end AI architecture supporting Level 4+, marking a key step from the traditional modular perception, planning, and control pipeline to a data-driven holistic neural network decision-making framework. However, TIER IV has not completely abandoned modularity; instead, it is advancing two routes simultaneously: a modular AI architecture and a holistic end-to-end architecture. The modular AI architecture introduces deep learning models into sub-modules such as perception, prediction, planning, and control, retaining the advantages of interpretability and safety isolation; the holistic E2E architecture maps sensor inputs directly to vehicle control outputs through neural networks.
The two routes are advancing in parallel, rather than being mutually exclusive. From the choices of these companies, a common point can actually be seen: the industry has not made a single choice between end-to-end and modularity. End-to-end models solve the problem of driving behavior, namely, how to make the car drive more like a human and require fewer takeovers in complex scenarios.
However, a complete autonomous driving system also needs to solve other problems: how to trace the cause when the model makes an error? How to safely degrade when sensors fail? How can regulatory authorities verify the safety of the system?
End-to-end models themselves cannot answer these questions. This is also why a layer of engineered safety fallback is universally retained in many mass-produced intelligent driving solutions. The reason Level 2 autonomous driving has been deployed on a large scale is that humans are still behind the steering wheel to provide the fallback; whereas for Level 3 and Level 4 to scale up relying purely on end-to-end approaches, breakthroughs are still needed in multiple aspects such as interpretability and safety verification.
03. Final Thoughts
So, by 2026, have traditional modules truly disappeared? Smart Driving Frontier believes that what has disappeared is the mechanical pipeline of perception, prediction, planning, and control, which is centered on the accumulation of manual rules. One-stage end-to-end is replacing these independent modules with a unified neural network.
However, the mindset emphasized by modularity—decomposing complex systems into independently verifiable and debuggable components—has not disappeared; it merely exists in a different way. It either transforms into observable component divisions within the end-to-end model, or into the rule layer outside the model used for safety verification, or into a dual-track strategy like TIER IV's, where modular AI and holistic E2E advance in parallel. Therefore, in 2026, what has disappeared in modularity is the rigid architecture, not modularity itself.
#Autonomous Driving #End-to-End Autonomous Driving #Autonomous Driving Modular Architecture