As autonomous driving enters 2026, the end-to-end (E2E) technical route has become a consensus, with nearly all leading players completing the transition to E2E architectures. However, the popularization of E2E has not brought about the anticipated technological leap. Data is exploding, fleet sizes are expanding, yet the pace of system improvement is slowing down. Many companies have entered a state where they appear to be progressing, but there has been no qualitative change in user experience. In fact, in the field of autonomous driving, the real competition after E2E has just begun.
01. How Far Can the Data Scaling Route Go?
E2E autonomous driving involves taking camera pixel inputs and outputting steering wheel angles, with no human-defined module boundaries in between. This architecture bypasses the issue of cascading errors in traditional modular architectures, but new problems arise. In many open-loop evaluations, there is a clear power-law relationship between model performance and data volume, but this rule does not hold in closed-loop evaluations that are closer to real-world driving. A study conducted both open-loop and closed-loop simulation tests on 1,400 complex driving tasks.
The results show that merely expanding data scale is insufficient to solve autonomous driving problems; the quality of data distribution and scenario coverage are equally critical. A small increase in long-tail scenario data can significantly improve the model's performance in corresponding complex scenarios.
Other studies have found inconsistencies in the data scaling laws in closed-loop evaluations, indicating that the focus of research in the autonomous driving field should shift from merely expanding data volume to data distribution. This is because E2E models are essentially imitation learning systems; they learn how humans drive, not why they drive that way.
When faced with long-tail scenarios that never appeared in the training data, its performance degrades sharply. In other words, it only knows how to handle what it has seen, not what it understands. Furthermore, vehicles generate massive amounts of video data daily, but the proportion of clips truly valuable for model training is extremely low.
A large amount of normal driving data not only contributes little but actually dilutes the model's learning ability for critical samples. The MOSAIC framework proposed by NVIDIA at CVPR 2026 offers a new solution. It first clusters and groups data by scenario type, then estimates the return curves for each data cluster across various driving metrics, and finally adds the data clusters into the training set one by one in descending order of marginal returns.
This mechanism prioritizes the selection of data that brings the maximum improvement under limited computational resources. Experimental results show that MOSAIC can achieve or even surpass the baseline performance of traditional full-data training with 80% less data. Besides data efficiency issues, E2E training in autonomous driving faces two other constraints.
Professor Li Shengbo from Tsinghua University pointed out at the High-Level Forum on the Development of Intelligent Electric Vehicles in April 2026 that, in addition to data filtering efficiency, insufficient data scale and limited computing power are also prominent bottlenecks. He provided a specific reference: the autonomous driving field requires approximately 100 million video clips to cross the entry threshold for model training, and a model parameter scale of 1B to 10B can basically support intelligent driving functions. Currently, the industry's data accumulation and model scale are still quite far from this threshold, leaving significant room for improvement.
02. From Seeing to Understanding: What Do VLA and World Models Bring?
Pure E2E models lack an understanding of the physical world. They do not know that a slippery road means an extended braking distance, that overhanging branches might fall, or that deep pits might be hidden under puddles. They can fit some superficial patterns through massive data but cannot establish causal reasoning capabilities. This is precisely why VLA (Vision-Language-Action) and world models have developed following the emergence of E2E.
VLA unifies vision, language, and action into the same model framework. The system directly generates driving actions from inputs such as video streams and navigation instructions, solving the problem of what to do. World models, on the other hand, learn the motion laws, spatial relationships, and causal structures of the traffic environment from massive videos by predicting the next frame and the state at the next moment, solving the problem of what will happen after taking a specific action.
Since E2E models lack an understanding of "why," world models can establish this understanding by predicting future states, while VLA translates this understanding into specific driving actions.
When the two are combined, the system can both understand the logic of how the world works and know what to do. Along this direction, multiple companies and research institutions have introduced new work in 2026, attempting to achieve world prediction and action decision-making simultaneously within a unified framework. The HyWorldVLA paper released by BYD Auto New Technology Research Institute is a typical representative.
The core innovation of this paper lies in the adoption of a hybrid world model, which uses both pixel-level and latent-space world models during the pre-training phase to leverage the strengths of both technical routes. Pixel-level prediction can provide fine-grained spatiotemporal reasoning but is sensitive to noise; latent-space models have strong anti-interference capabilities but limited interpretability. HyWorldVLA attempts to find a balance between the two. The model achieved a PDMS score of 90.59 on the NAVSIM v1 benchmark, reaching a leading level among publicly available results.
Image source: Internet. In addition, there is much other similar work. WCogVLA proposes a dual-level world cognitive VLA framework, combining semantic world prediction with generative world evolution. CoWorldVLA is a multi-expert world reasoning framework that uses world representations as explicit conditions to guide action planning. The commonality of these works is that they no longer treat VLA and world models as two independent technical routes but attempt to closely integrate them at the representation level.
03. Data Closed-Loop: From One-Time Training to Continuous Evolution
If VLA and world models solve the problem of model comprehension capabilities, then the data closed-loop solves the problem of model evolution capabilities.
In the past, the process from data collection to model deployment was linear. Data was collected at the vehicle end, manually annotated, used to train the model, and finally deployed on vehicles. The cycle of this entire chain was very long, taking days or even longer from collection to deployment. After the fleet size expanded, although the data volume increased, the bottleneck in processing capacity further amplified this delay, ultimately resulting in an accumulation of more and more data while the model's cognition updated very slowly.
The new approach now is to turn the linear process into a cycle. Vehicles collect data while driving, and the system automatically mines noteworthy difficult scenarios from it. These scenarios can be directly used for model training. After training is completed, they are pushed to the fleet via OTA (Over-The-Air). The fleet hits the road with the new model and continues to collect new data. Every time the cycle turns, the system can evolve once.
The efficiency of this closed-loop depends on the degree of automation in each link. Whether the collection link can perform preliminary screening at the vehicle end to reduce the upload of useless data; whether the mining link can accurately identify scenarios that bring the most improvement to the model; whether the training link can complete iterations quickly; and whether the OTA link can be deployed safely and efficiently. These links are connected in series; if any one gets stuck, the entire cycle will slow down. We can hear the term "data closed-loop capability" in many places, which refers to making this cycle run fast enough and stably enough.
In the data mining link, Mobileye introduced two AI tools, Meteor and Genario, at CVPR 2026. Meteor is a hypothesis-driven data mining engine, and Genario is a targeted scenario simulator. Both aim to identify and solve long-tail problems in autonomous driving through systematic methods.
Meteor is a multi-agent AI data analysis platform built by Mobileye specifically for autonomous driving, aggregating millions of hours of driving data from different countries, weather conditions, road types, and traffic environments. It can automatically identify reproducible faults, deduce causal hypotheses, retrieve similar scenarios, and generate training samples, operating at scale with full automation throughout the process.
Genario, based on Meteor's output, generates photorealistic synthetic driving scenarios, expanding the coverage of targeted training data by switching controlled variables such as lighting, weather, and obstacle positions. The core idea of these tools is not to passively wait for long-tail scenarios to appear, but to proactively mine those safety-critical rare cases from massive data. In the annotation link, the combination of AI pre-annotation and manual refinement is changing traditional data production methods.
The integration of unsupervised, weakly supervised, and few-shot supervised technologies is creating a semi-automatic annotation mode of zero, minimal, and precise annotation, which can improve efficiency by over 70% compared to traditional manual annotation. In the iteration link, OTA has become a standard feature. When the degree of automation in each link of this pipeline is high enough, it is no longer just a project but a system capable of self-iteration and continuous evolution.
04. Final Thoughts
The competition in E2E autonomous driving is shifting from an algorithm race to a system operation race. In the past, the competition was about who could design a better network structure and who could run through E2E first. Now, it is about whose data closed-loop turns faster, who can mine long-tail scenarios more accurately, and who can go further in the integration of VLA and world models.
The technological gap among leading companies is fluctuating on a monthly basis, and no one can establish a long-term advantage through a single technological breakthrough. For these companies, the true barrier lies not in a single model, but in a complete closed-loop system from data collection, scenario mining, model training to OTA deployment. The operational efficiency of this system determines the speed of model evolution.
Although data scale remains important, how to use data well is more critical than how much data there is. VLA and world models provide the possibility for models to understand the world, but the prerequisite is having enough high-quality data and a highly efficient closed-loop to support their continuous evolution.
#AutonomousDriving #EndToEndAutonomousDriving #WorldModels