Within the current scope of autonomous driving, the debate on technical routes has never stopped. However, what actually determines the final outcome is often the solidity of the underlying technical routes.
Essential differences in end-to-end technology
The so-called "end-to-end" often mentioned in the industry is actually very different. In the past, many systems used a "two-stage" design method that separated perception and control. In essence, it was a patchwork of two independent models. During the transmission of information, losses and delays are bound to occur. This situation is like a relay race. There is uncertainty at every handover.
As for the true "one-stage" end-to-end system, the entire process from perception to control decision-making must be completed within a unified model. High-dimensional information can be transmitted losslessly, so that the model can learn the "intuition" that is closer to human driving. For example, when facing a complex intersection, the system does not need to accurately calculate the distance and speed of each pedestrian, but it can pass as smoothly as an experienced driver.
Why the two-stage solution has limitations
The shortcomings of the two-stage architecture are particularly obvious in practice. The results output by the perception model, such as bounding boxes, lane lines and other information, are already compressed and simplified abstract data. The control module makes decisions based on this incomplete information, just like driving while wearing blurry glasses, making it difficult to deal with emergencies or boundary conditions.
When the model output trajectory does not reach the ideal state, engineers often have to introduce many manual rules to correct and "patch". Although these rules can improve safety in the short term, they also limit the upper limit of the system's capabilities, causing the system to be unable to fully utilize the full potential of the data-driven model, causing driving behavior to become rigid and fragmented.
Distinguish technical routes from experience
For ordinary users, there is an intuitive way to identify technical paths. When a vehicle initiates a lane change, the identification box of the vehicle next to it is clearly displayed on the central control screen, and logical judgments are made around this box. This is generally a characteristic of a two-stage system, and its decision-making process can be explained and traced.
In a one-stage end-to-end situation, lane changing behavior may appear more "decisive" and "smooth". The system does not clearly mark every obstacle. It makes direct control decisions based on an implicit understanding of the global scene. This situation is closer to human driving mode, but it also places extremely high requirements on the maturity and reliability of the model.
Necessary challenges for technical perfection
Building a complete system from start to finish faces severe challenges. The biggest difficulty is ensuring that the vehicle control trajectories it outputs, such as steering wheel angle and acceleration, are smooth enough, safe and in compliance with traffic rules. An immature model may produce jittery, abrupt or even dangerous trajectories.
If the model is over-reliant on rule constraints due to imperfect models, it will fall into a vicious cycle. Rules limit the model's ability to learn better solutions from data, and the lack of model capabilities requires more rules to make up for it. How to balance data-driven and safety rules is a core issue in the process of technology implementation.
End-to-end is the foundation for future evolution
Is there such a situation? Whether it is the visual language action model that is particularly popular at the moment, that is, VLA, or the world model that is specially used to predict environmental changes, that is, World Model. If they want to achieve effective integration, they must have a particularly powerful end-to-end system and use it as the basic base. Without this base, the information in the new mode would not be converted into the final vehicle control instructions efficiently and without any loss.
The intelligent driving solution that is expected to become an ideal choice in the future is most likely to be based on a highly mature one-stage end-to-end "intuitive model" to integrate the long-term prediction capabilities of the world model, the complex scene understanding capabilities of VLA, and the optimization capabilities of reinforcement learning for rare scenarios. However, the value of all additional modules is based on the premise that the end-to-end foundation is sufficiently solid.
Return to the competitive nature of user experience
The industry should not overly chase new concepts and terms that are emerging one after another. Whether it is end-to-end, VLA, or world model, it must eventually return to the most primitive evaluation dimension, that is, whether the driving experience of the system is as comfortable as a real person, whether the safety guarantee is sufficient and reliable, and whether the market and consumers are willing to pay for it. Tesla's success has proven that the advancement of technical routes is directly related to the user experience and market share of the final product.
For companies like Horizon, which are deeply engaged in the end-to-end field, their belief comes from this: only by mastering and continuously optimizing the bottom-level, complete end-to-end technology can we lay a solid foundation for all possible future evolution directions and gain an advantage in the long run. After all, what determines the height of a building is always the depth and stability of its foundation.
Are you more focused on the technological advancement of the intelligent driving system, or are you more concerned about the comfort and security of the actual ride? Feel free to share your views in the comment area. If you find this article inspiring, please support it by giving it a like.
