Currently, among the high-end intelligent driving systems on the market, many are still in a "usable" state. The distance can make users feel "easy to use" and actively express their "love to use". There is a long and far road waiting to be crossed. The most critical obstacle here is that the traditional rule-based technical approach faces great difficulties in effectively dealing with the infinitely complex and diverse driving scenarios in the real world.
Limitations of traditional rule systems
Traditional autonomous driving systems generally adopt a modular design, with the sensing link running independently, the prediction link running independently, the planning link running independently, and the control link running independently. This architecture relies heavily on many "if-then" rules written by engineers. In the high-speed pilot assist function that will be mass-produced around 2020, such rules can also deal with structured roads.
But when the system enters complex urban roads and faces random pedestrians, non-motorized vehicles, and unruly vehicles, the preset rules are stretched. System behavior often appears rigid, system behavior often appears conservative, frequent unnecessary stops occur, and frequent "dareness" to change lanes occurs, resulting in low traffic efficiency and greatly reducing the user experience.
The core concept of end-to-end technology
There is a disruptive idea called end-to-end technology. It does not divide the driving task into multiple modules, but tries to build a unified deep neural network model. This model takes the raw data from the vehicle sensors, such as images and lidar point clouds, and directly maps it to the final driving operation instructions.
Such an integrated design prevents the loss of information transmission between modules and the accumulation of errors. From a theoretical level, the model can learn a decision-making method closer to that of real drivers from massive human driving data, covering the understanding and processing of fuzzy scenes.
A key breakthrough from idea to mass production
Transforming the end-to-end concept into a stable product that can be mass-produced will encounter great challenges. The core difficulty lies in how to ensure that the model is safe and interpretable. The early end-to-end model is like a "black box", and its decision-making logic is difficult for engineers to understand and intervene.
Over the years, with the continuous advancement of related technologies called "large models" and the methods used in simulation testing, this difficult problem is in the process of being overcome. By constructing a large number of extreme scenarios in the simulation environment for testing, and introducing new explanatory model tools, those engaged in R&D can deeply understand the decision-making basis of the model, and thus improve performance while firmly adhering to the lowest limit of safety.
Improve complex scene interaction capabilities
The significant advantages of the end-to-end system are presented in the interaction and gaming capabilities for complex dynamic scenarios. For example, when it is time to merge into the main traffic flow on a congested road section, the traditional system may wait for a long time because it cannot predict the reaction of the nearby vehicle.
However, the data-driven end-to-end model, by learning the game data of real human drivers, can more accurately pre-judge the intentions of surrounding vehicles, and then make bolder and smoother lane-changing decisions. Test data shows that in some complex intersection scenarios, the traffic efficiency of the system based on new technologies can be improved by more than 60%.
Reduce dependence on high-precision maps
In the large-scale promotion of smart driving, over-reliance on high-precision maps has always been a stumbling block. The cost of producing high-precision maps is extremely high, and updates are not timely, making it difficult to fully cover every city and rural road in the country. End-to-end technology has opened up a new path to resolve this problem.
Relying on the real-time mapping and positioning capabilities of the enhanced perception model, the system can understand the current road geometry, lane lines, and traffic signs just like a human driver, mainly using its own "eyes" and "brain". This enables high-end intelligent driving functions to be implemented more quickly in areas without images.
Accelerate the popularization of smart driving in all scenarios
Gradually maturing end-to-end technology will effectively promote high-end smart driving to expand from high-speed scenarios to all scenarios such as urban areas and parking. A software and hardware solution with powerful performance and controllable cost is the key to popularization. It can help car companies launch smart driving products with a better experience with shorter development cycles and lower costs.
From the perspective of consumers, this shows that in the next few years from now, we hope to drive cars that can provide reliable and smooth driving assistance functions in most daily commuting situations, and truly transform from "occasionally used" to "relied on every day."
Let me ask you, in your opinion, a smart driving car that can really make you feel "easy to use" or even "love to use", in addition to bold lane changes and efficient traffic, in which daily scenes must it also show outstanding performance? You are welcome to submit your opinions in the comment area. If you agree with the views of this article, please like it and support it.
