The Beijing Auto Show continues.
Like the Shanghai and Guangzhou Auto Shows in 2023, intelligent driving is still the unavoidable theme of this auto show, but at the same time, the market is also changing rapidly, and the speed of technology iteration is exceeding everyone's expectations.
"From last year's Shanghai Auto Show to this year's Beijing Auto Show, the development of intelligent driving has advanced by leaps and bounds, and even I was amazed. ”
Dr. Lang Xianpeng, Vice President of Intelligent Driving of Li Auto, said at the Horizon Ecological Roundtable Forum during the Beijing Auto Show, "Last year, everyone was still talking about high-speed NOA, Transformer and BEV, but this year the industry has become accustomed to urban NOA, and almost all of them are talking about mapless solutions and end-to-end. ”
In order to improve the market penetration rate of L2+ intelligent driving functions, the industry is now competing for the large-scale competition of urban NOA, which is shifting from the speed of opening the city to the daily activity.
The daily activity will become a key indicator to measure whether the NOA in the urban area is easy to use, and consumers' actual experience evaluation and preference for the intelligent driving function are also being put in an increasingly important position by various car companies.
However, one problem is that, from the actual experience, the urban NOA has not yet reached the "usable" standard in the eyes of users, and it is far from "loving to use".
"It's very difficult for urban NOAs to really make consumers want to use them. At present, the activation time of NOA in urban areas is actually very low. Zhou Guang, CEO of Yuanrong Qixing, said, "Even if you are willing to use it, many times, these functions bring fright, not surprise, to consumers." It may be once (an experience that doesn't go well) that will make the consumer quit. ”
From "usable" to "easy to use", how long will it take for high-end intelligent driving? Will the end-to-end that many car companies and Tier1s almost all talk about be the answer to this turning point? What will be the best solution for end-to-end mass production?
High-end intelligent driving "Kaicheng War" involution accelerates: user experience is still a "pain point"
The gunpowder smell of various car companies about high-end intelligent driving "Kaicheng War" has spread all the way to the Beijing Auto Show.
"We are definitely the first to open urban NOA in the whole country, and our goal is to open it in most parts of the country and even in small areas by the end of this year. He Xiaopeng, CEO of Xpeng Motors, said in an interview with Xinzhijia and other media at the Beijing Auto Show.
Lei Jun confidently shouted at the auto show, "The high-speed NOA of Xiaomi SU7 is definitely comparable to Huawei and Xiaopeng." At the same time, Lei Jun also announced that the city NOA of Xiaomi SU7 will be pushed in May.
At the same time, Yangwang Automobile's Yangwang U7 will open the city navigation function in the third quarter of this year, the Blue Mountain Intelligent Driving Edition of the Wei brand will conduct a large-scale NOA test drive experience in May, and Jiyue Automobile plans to achieve urban NOA coverage of the whole country this year......
From the perspective of time, a number of domestic new energy brands are entering a new stage of large-scale mass production of high-end intelligent driving, which can be opened nationwide and "can be opened when there is a road".
However, in fact, in the past two years, car companies have encountered considerable challenges in the large-scale implementation of high-end intelligent driving solutions.
In the mid-to-high-end intelligent driving solution, due to the immaturity of the technology, the functions that should bring consumers the ultimate experience cannot play their due role, so as to attract consumers to pay.
McKinsey pointed out in the "2024 McKinsey China Automotive Consumer Insights" report that consumers' interest in various autonomous driving functions is increasing, but the willingness to pay extra for them is declining, especially in first-tier cities.
In fact, although the NOA function in urban areas has been on the car for a long time, and various car companies also attach great importance to the landing and promotion of the NOA function in urban areas, from the perspective of market feedback, most car companies' urban NOA is still in its infancy, and it is still unable to meet the user's intelligent driving needs.
At the Beijing Auto Show, Horizon pointed out that the intelligent driving system will go through three stages of development from usable, easy to use to easy to use:
•Available: With the goal of meeting physical indicators, we focus on physical indicators such as scene pass rate and commuting efficiency, but in terms of experience, they are either "cowardly" or "reckless", and can only meet the standard of "usable".
•Easy to use: On the basis of physical indicators, with the goal of achieving an anthropomorphic experience, we focus on achieving a more elegant, calm and ready-to-activate experience, reaching the standard of "easy to use" in the hearts of users, and truly changing consumers' cognition.
•Love to use: With the goal of achieving equal rights for intelligent driving, we focus on making high-end intelligent driving experience accessible to everyone through the ultimate application and mass production efficiency, and truly let every consumer "love to use".
At present, the high-end intelligent driving system represented by urban NOA has not yet fully reached the "usable" state in the eyes of users, and the current four major problems of urban NOA, such as "cowardice", "recklessness", "urgency" and "expensive", have become the key factors that hinder consumers from paying:
1. Instigation: Congestion adds to congestion, leading to experience taking over;
2. Recklessness: The rush is too reckless, resulting in a safe takeover;
3. Urgent: In order to quickly open the city, the availability of the system decreases;
4. Expensive: It is not conducive to functional inclusion and large-scale market landing.
Taking MPI as an example, at present, when users turn on the NOA function in urban areas, they basically have to take over manually once for tens of kilometers, which is far from achieving "easy and easy to use", taking typical high-frequency scenarios such as turning at intersections as an example, at present, users usually choose to take over because of the system fullstop or stiff trajectory when turning on the NOA function in urban areas, which is far from achieving "easy and easy to use".
Coupled with the poor user experience, the high cost, and the limitation of high-precision map coverage in all scenarios, the NOA function in urban areas can only be opened in a limited number of high-tier cities, and the overly aggressive removal of high-precision maps also brings a certain degree of system performance degradation, and various de-function problems often occur randomly, making the actual user frequency of these functions not high, which further causes the problem of difficult iterative development of functions and long cycles.
In order to further increase the market penetration rate of intelligent driving, in 2024, the improvement and iteration of intelligent driving functions must shift to practicality and scenario optimization, with "user stickiness" as its core competitive point.
At a time when the price war is becoming more and more intense, how to achieve a better functional experience at a limited cost, and how to make the next generation of high-end intelligent driving systems take into account both traffic efficiency and anthropomorphic experience?
Horizon believes that technology reconstruction is the only way to achieve mass production of high-end intelligent driving systems in all scenarios.
In 2023, Tesla announced the FSD V12 version, upgrading the urban street driving stack to the end-to-end neural network technology route, so a wave called "end-to-end" craze began to sweep the domestic intelligent driving circle, and at this year's Beijing Auto Show, it is gradually heating up, becoming one of the keywords that car companies and Tier1 can't put down.
The leap from "usable" to "easy": open the next shuffle end-to-end
The so-called end-to-end refers to a new type of AI model, which uses BEV+Transformer technology architecture and other methods to realize the integration of perception and decision-making, so as to achieve the effect of outputting the final execution instructions after inputting the original data. This technology reduces the number of codes by hundreds of thousands of codes compared to previous versions, allowing the car to drive on unfamiliar terrain without a data connection.
An algorithm solution development engineer once pointed out to the new intelligent driving, "The integration of perception and decision-making into the same model makes the end-to-end model effectively avoid the error value between the levels, without the intervention of any manual rules, and is closer to the high-level intelligent driving behavior of human driving." ”
In 2023, after Tesla's FSD V12 demonstrated the effect of the end-to-end model on the car, domestic mainstream car companies and Tier1 realized that this was a future trend and began to catch up quickly.
In the past, most intelligent driving practitioners divided the entire autonomous driving task into modules such as perception, prediction, decision-making, and control, and each engineer was independently responsible for one or two of the modules because the technology stack of each module was very different and difficult.
Therefore, most of the intelligent driving models that have been mass-produced at present also adopt the traditional modular architecture, that is, they are divided into different small models according to the functions of perception and prediction, and each model must be trained and optimized separately, and the downstream regulation and control links are still dominated by rules.
Different from the traditional modular architecture, the end-to-end model is cascaded by multiple small models (i.e., neural networks), and only one large model needs to be trained to optimize and improve the capabilities of each functional module, thereby reducing the R&D costs caused by module-by-module training under the traditional architecture.
(New Intelligent Driving Finishing and Drawing)
On the eve of the outbreak of the technology landing war, the first step is academic pre-research.
Academic end-to-end research began with ALVINN in 1988, followed by the development of an end-to-end CNN prototype system.
As early as 2016, Horizon took the lead in proposing the concept of end-to-end evolution of autonomous driving, and in 2017, it began to adopt the end-to-end system of training, and at the same time released the software framework Hugo.
In 2022, Horizon Robotics proposed Sparse4D, an industry-leading end-to-end algorithm for autonomous driving perception, and in 2023, UniAD, the industry's first publicly published end-to-end autonomous driving model by Horizon Robotics scholars, won the Best Paper in CVPR 2023.
In this paper, the UniAD framework was proposed for the first time, becoming the industry's first general model for autonomous driving with the integration of perception and decision-making.
Based on this, the researchers integrated three main tasks such as perception, prediction and planning, and six sub-tasks (object detection, target tracking, scene mapping, trajectory prediction, raster prediction and path planning) into a unified Transformer-based end-to-end network framework to realize a full-stack mission-critical driving general model.
At the same time, Horizon has also accumulated an end-to-end deep learning algorithm based on Monte Carlo tree search, which has greatly improved the perception and interactive game ability of the intelligent driving system in complex traffic environments.
An industry executive revealed to Xinzhijia, "Strong companies are expected to launch end-to-end models within 1-2 years, and slowly, suppliers and regular enterprises will follow, which is a definite trend." ”
For example, in December 2023, Li Auto completed the OTA 5.0 update. According to the release of AD Max 3.0, with the support of the end-to-end architecture, Li Auto integrates capabilities such as BEV large model, MPC model predictive control, and spatio-temporal joint planning, and also adds the occupancy network algorithm. At the same time, Ideal also uses the self-developed neural prior network NPN (NeuralPriorNet) to "patch" BEV.
In addition, in order to deal with complex traffic light information, Li Auto has said that it uses Traffic Intention Net (TIN) to solve the problem, and TIN is an end-to-end model.
He Xiaopeng also revealed at the Beijing Auto Show that Xpeng Motors will launch an end-to-end large model AI intelligent driving in the second quarter, and the application of this large model will be converted from the past image-based algorithm framework to video streaming.
As companies race to get the end-to-end model into the car, the industry has been repeatedly asked whether the end-to-end model will become the technological endgame of high-end intelligent driving.
Industry insiders told the new intelligent driving that the ultimate development direction of intelligent driving technology should be closer to the behavior pattern of human driving, at least with a stronger and closer relationship with modules similar to human driving behavior.
In his view, the final technical architecture of high-end intelligent driving should be a completely end-to-end solution, with cameras, millimeter-wave radar, lidar and other multi-sensor, multi-modal inputs, as the end-to-end input, the vehicle's control command becomes the output, skipping the middle positioning, prediction, planning, decision-making and other modules, only the input and output of these two ends, which is the future technical framework.
Taken together, an end-to-end approach means that larger parts of the system are built with data-driven modules, which means that a corresponding proportion of modules are maintained by humans, making the overall system easier to maintain.
In addition, the end-to-end architecture realizes the functions of multiple models through one model, and R&D personnel only need to train, adjust and optimize the overall model to achieve performance improvement, so they can better concentrate resources and achieve function focus.
More critically, the potential of the end-to-end model approach lies in having a higher performance ceiling. End-to-end autonomous driving can improve computational efficiency by reducing tasks and avoiding a lot of repetitive processing, and can push the limits of a system's capabilities by continuously expanding data.
Functionally focused, easy to maintain, with higher performance ceilings, closer to human driving behavior...... The end-to-end system is undoubtedly the optimal implementation path to drive the urban NOA towards "easy to use and love to use".
With the acceleration of the implementation of the end-to-end system, the competition of the industry on high-end intelligent driving will also usher in a new node.
In fact, the current bottleneck in the development of end-to-end autonomous driving technology is still the traditional three major problems - algorithms, computing power and data.
Among them, the computing platform needs to support the network in the larger-scale end-to-end autonomous driving system, the data layer needs to have massive and very diverse data that can cover most cases in the life scenarios of ordinary users, and the algorithm needs to explore an interpretable and intervenable end-to-end autonomous driving system.
Whoever can take the lead in breaking through these three levels will undoubtedly have the confidence to go to the end and the decisive hand in this end-to-end shuffle.
The road to end-to-end mass production: software and hardware synergy is the best solution
This means that players in various industries must solve the dual challenges of complex embedded computer system technology and extreme engineering capabilities to achieve efficient collaboration between software and hardware.
This aspect requires companies to continuously expand ODD, and to achieve the highest possible product performance in as large a geographical space as possible and in all weather conditions.
On the other hand, it is necessary to improve the performance and experience of functions, realize NCA, ICA+, ICA, VPA, APA and other working modes and switch them seamlessly, and polish the system performance of each mode to the best to improve the system's ability to handle hard cases.
At present, players in the industry usually have two types of technical strategies when landing high-end intelligent driving.
The first is to focus on improving the functional experience and performance, but the expansion of ODD is not enough, and it is still very narrow, and the second is to focus on the expansion of the broader ODD, but the level of functionality is very narrow and limited.
To achieve both, developers need to have strong software and hardware capabilities, as well as extreme engineering capabilities, to achieve breakthroughs in both dimensions.
In fact, hardware configurations such as higher-end computing solutions and sensors determine the lower limit of performance, and more advanced software technology architecture determines the upper limit of performance.
In the view of Yu Kai, the founder of Horizon, the collaboration of software and hardware is the unique advantage of Horizon, relying on the leading cutting-edge software algorithm research to design the most advanced hardware computing architecture, and then support the most advanced algorithms, so that the chip architecture and application software are closely matched, which can make the computing scheme very efficient.
At this Beijing Auto Show, with cutting-edge end-to-end software algorithm technology, Horizon released SuperDrive, a mass-produced high-end intelligent driving system that combines end-to-end technology, in an attempt to provide a solution to create an "easy-to-use" intelligent driving solution.
On the one hand, SuperDrive increases the occlusion call rate of the intelligent driving system by 70% through the three-in-one sensing end-to-end architecture, effectively solving the problems of high latency, multiple rules, and heavy load in the current industry perception architecture.
On the other hand, SuperDrive has greatly improved the perception and interactive game capabilities of the intelligent driving system in complex traffic environments through data-driven interactive games, increasing the success rate of lane change by 50% and the traffic efficiency of intersections by 67% in congestion scenarios.
At the same time, SuperDrive also has the ability to perceive without high-precision maps, and can achieve light dependence on high-precision maps, so as to quickly launch the NOA function of each city and achieve efficient city expansion.
At the level of computing solutions, SuperDrive collaborates with the Journey 6 flagship version to create the best high-end intelligent driving system that combines software and hardware, so that the easy-to-use urban NOA solution can accelerate the large-scale mass production and achieve accessibility for everyone.
In fact, in addition to SuperDrive, an urban NOA solution, at the Beijing Auto Show, Horizon also released the industry's long-awaited Journey 6 intelligent driving computing solution.
The Journey 6 series has launched a total of six versions, including Journey 6B, Journey 6L, Journey 6E, Journey 6M, Journey 6H, and Journey 6P, among which the Journey 6 flagship version - Journey 6P, with a computing power of 560 TOPS, is oriented to the high-end intelligent driving market, which has six product characteristics, such as high integration, high computing power, high efficiency, high processing power, high access capacity and high security.
According to the introduction, a single Journey 6 flagship can support full-stack computing tasks such as perception, planning and decision-making, and control, and support full-scene NOA functions.
At the same time, Journey 6P also achieves high cost performance under the premise of high performance, and the cost of hardware system based on Journey 6P can be less than 10,000 yuan.
On the other hand, Transformer is also an important technology for realizing end-to-end autonomous driving.
If end-to-end is compared to the end point of high-level intelligent driving, BEV+Transformer is more like a bridge in the journey, and can also be regarded as a tool, while the occupancy network is an auxiliary "weapon" that can be used superimposed.
Horizon has achieved the best support for Transformer in terms of algorithm and computing scheme.
At the algorithm level, UniAD is the industry's first all-in-one end-to-end model for perception and decision-making, achieving the best performance on all tasks in nuScenes.
Horizon's pure vision autonomous driving algorithm Sparse4D is its next-generation architecture that takes over from BEV+Transformer. In January this year, Horizon also open-sourced the Sparse4D series of algorithms to promote more developers in the industry to participate in the exploration of cutting-edge technologies such as end-to-end autonomous driving and sparse perception.
At the hardware level, Horizon's next-generation BPU Nash architecture natively supports Transformer models with large parameter quantities and provides the strongest Transformer computing performance in the industry, providing strong support for the deployment of advanced end-to-end systems.
In terms of related technology stacks, the various tools built by Horizon to promote software development and customization, such as algorithm development tool chain Tiangong Kaiwu, embedded middleware stepping song, software development platform Eddie, etc., also reflect its strong software and hardware comprehensive capabilities.
The full-stack software and hardware capabilities with lower cost, higher performance ceiling, and more efficient collaboration point to an ultimate direction for the collaboration of the high-end intelligent driving system SuperDrive+ Journey 6 Ultimate Edition - to help industry players more efficiently achieve large-scale mass production of high-end intelligent driving based on end-to-end models.
According to Horizon, in terms of cooperative car companies, Horizon's SuperDrive solution has reached cooperation with a number of car companies and top Tier 1, and it is expected that by the fourth quarter of this year, there will be a standard mass production plan launched, and by the third quarter of next year, the first model equipped with SuperDrive solution will also enter the delivery period.
From usable to easy to use, the road of high-end intelligent driving in China has just reached a turning point, and in order to win in the hot battle on end-to-end, as the head supplier of domestic intelligent driving solutions, the journey of Horizon has just begun.
Leifeng.com#Leifeng.com#Leifeng.com