Author: Tian Xi
In 2022, driven by multiple factors, intelligent driving systems are really booming.
According to data from the Ministry of Industry and Information Technology, the penetration rate of L2-level assisted driving passenger cars in the new car market reached 30% in the first half of this year, an increase of 12.7 percentage points year-on-year.
In terms of specific automakers, the proportion is even more prominent:
Taking the data of Great Wall Motor in August as an example, a total of 88,226 new cars were sold, of which the proportion of intelligent models increased to 85.59%, a historic new high.
In the past, intelligent driving was considered the representative of new forces in the automotive industry such as Tesla and WeRide, but now it is gradually becoming standard equipment for cars.
The reason behind this change is not difficult to understand.
As the intelligentization of cars becomes a consensus, more and more traditional car companies are working with autonomous driving companies such as Baidu, Momenta, Huawei, etc. to develop products with L2/L2+ functions to adapt to the competition in the new era through self-research or cooperation.
Breakthroughs in the regulatory field have also provided impetus for the development of intelligent vehicles.
In July of this year, Shenzhen passed legislation for the first time to open the access and registration of intelligent connected vehicles to clear obstacles for the large-scale production and landing of the latter.
Momenta CEO Gu Weihao believes that the data scale needs to be large enough, and the mileage of autonomous driving needs to reach at least 100 million kilometers. In addition, data diversity also needs to be sufficiently sufficient.
Tesla is undoubtedly leading the way, and its first “data closed loop” gyroscope is turning faster and faster.
According to Tesla’s 2022 Q2 financial report, FSD Beta version test users have accumulated about 42 million miles of driving, and this is only based on the foundation of 100,000 test users.
According to Musk’s plan, if the number of test users expands to 1 million by the end of this year, the cumulative driving mileage of FSD Beta will soon exceed 100 million miles.
In China, the growth rate of the mass-produced assisted driving scale and the speed of autonomous driving technology evolution can only be compared with Tesla by Momenta, perhaps the only one.
How to obtain high-quality and high-value data of vehicle travel on a large scale at low cost and high efficiency, and use it to iterate technology, has become the focus of competition among various players.
On September 13th, at the 6th HAOMO AI DAY of Haomo Zhixing, Zhang Kai, the chairman of Haomo Zhixing, revealed that the cumulative driving mileage of its assisted driving service has exceeded 17 million kilometers, ranking first among Chinese self-driving companies.
It is worth mentioning that this milestone was achieved exactly 1,000 days after the establishment of Haomo Zhixing.
Through the Storm and Stress of 1,000 Days, Haomo Zhixing Overcomes Three Major Difficulties: Technology, Mass Production, and Commercialization
As a startup company, how did Haomo Zhixing become the first to break through in the highly competitive autonomous driving industry and achieve the title of “the company with the highest domestic operational mileage”?
The answer provided by Zhang Kai is Haomo Zhixing’s unique “Triangle of Intelligent Driving Product Capability Iteration” methodology.
“We believe that autonomous driving relies on highly effective collaboration among three aspects: scenario-based user experience design, artificial intelligence technology, and engineering and technical capabilities,” he said.
In terms of scenario-based user experience design, Haomo Zhixing has developed an integrated design method for user interaction experience and product development based on the production experience of multiple car models. By continually updating and iterating user tracking data, Haomo Zhixing is able to fully understand user usage habits.
This makes Haomo Zhixing the first company in China to use real user data to drive product iteration.
In terms of artificial intelligence technology, Haomo Zhixing launched MANA, China’s first intelligent driving data intelligent system, based on the most advanced AI technology concept in the world in December 2021, which became the core driving force behind the iteration of all Haomo Zhixing products.
In terms of engineering and technical capabilities, Haomo Zhixing developed three generations of intelligent driving systems in just two years, achieved mass production and landing of more than 10 different platform vehicles, realized the reuse of new vehicle models, and achieved mass production and landing in just four months.
In summary, scenario-based user experience design is the entry point, artificial intelligence technology is the soul, and engineering and technical capabilities are the guarantee.
These three capabilities positively interact and support each other, allowing Haomo Zhixing’s intelligent driving product capabilities to be rapidly iterated.
With the “Triangle of Intelligent Driving Product Capability Iteration”, Haomo Zhixing overcame the three major difficulties in the industry from 0 to 1 in the past 1,000 days: mass production difficulties of large-scale and multi-type autonomous driving, cost difficulties of terminal logistics automatic distribution vehicles, and difficulties in data processing and large model applications on a large scale.“`markdown
Using the challenge of mass production as an example, since the mass production of HPilot1.0 in May 2021, with the Wei Pai Mocha being the first model equipped with the HWA high-speed assisted driving system, in April 2022, HPilot2.0 was mass-produced, using the sixth model of automotive driving assistance system by Mobileye, the Tank 500. For two and a half years, Mobileye has stably delivered three generations of driving assistance products for passenger cars, covering more than ten models.
Now, HPilot welcomes version 3.0 and will officially land within 2022, becoming the first Chinese company to truly mass-produce a high-level NOH urban assisted driving product.
In terms of end-to-end logistics automatic distribution, Mobileye occupies an absolute market-leading share, with the launch of the Little Camel 2.0 mass production and delivery to customers.
In addition, Mobileye has created the first data intelligence system in China, MANA, which completes the labeling of hundreds of thousands of full elements and multi-modal CLIPS, and accumulates a 3 million-hour cognition scene library for driving on Chinese roads. Then, after massive training and learning, the virtual driving experience is equivalent to that of human drivers with 40,000 years of experience, essentially completing a data loop.
“What is unique about Mobileye?”
Zhang Kai believes that the “iron triangle” is precisely the advantage that sets Mobileye apart from other autonomous driving companies since its founding.
Thus, the judgment on the path of intelligent driving is the “worldview” of the company.
Only with the correct worldview can the methodology be used successfully.
The autonomous driving industry has always had a debate between the “gradual” and “leapfrog” routes.
Zhang Kai stated that Mobileye’s worldview is based on the fact that it has always decided to take a gradual development route.
And at this year’s AIDAY, Zhang Kai clearly stated: “Assisted driving is the only way to fully autonomous driving.”
This statement is likely to spark further heated industry discussions.
In Zhang Kai’s view, the gradual route has a much earlier production time and can quickly form large-scale operations, accumulating enough real user data from usage scenarios.
Moreover, compared to the directed data collection approach of the leapfrog route, the gradual route is more cost-effective and of better quality.
“When we combine the automatic driving product capability curve we have summarized, we can see that the scale and cost of data, as well as data quality, are directly correlated with the speed of automatic driving product capability enhancement.”
In fact, Zhang Kai’s statement touches on the current development trend of the industry.
In 2022, more and more companies are choosing to start from L2, even including L4 companies that were once steadfast in their “one-step approach.” These companies are seeing that technology landing is still far off, and have turned to cooperating with car companies to seek mass production.
It is clear that intelligent driving has quietly entered the second half. At the same time, new challenges arise.
“`For Huan Mo, the latest three challenges for large-scale autonomous driving data migration to the cloud, breakthroughs in AI chip performance, and mass production of urban scene-assisted driving products constitute the latest three challenges.
To this end, Zhang Kai proposed the “Five Winning Rules for Winning the Second Half of Intelligent Driving”:
- The development of intelligent driving products always puts safety first;
- The product experience should be “delicious”;
- Driven by real user scenario data, achieve rapid product iteration;
- Achieve highly integrated perceptual and cognitive intelligence;
- Empower customers with an open mentality to promote shared progress in the industry.
Zhang Kai particularly emphasizes the principle of safety first, and as intelligent driving faces increasingly complex scenarios, safety becomes more important, and in fact, this is also the basic starting point for the development of all intelligent driving products.
In terms of product experience, Huan Mo believes that only by using To C thinking to do To B things can truly develop products that can be accepted by the C end market.
Data-driven is Huan Mo’s core principle, and Zhang Kai said that by mining real user usage scenario data to achieve rapid product iteration and thus improve user interaction perception.
In addition, opening up perceptual and cognitive intelligence and empowering customers with an open mentality to promote shared industry development will also be the focus of Huan Mo’s efforts to come.
Zhang Kai summed up: “We are overcoming the challenges of the epidemic, supply chains, and technology itself, and will strive to achieve our mass production and delivery targets set earlier this year in the next four months.”
In the era of Autonomous Driving 3.0, data intelligence wins
At this year’s HAOMO AI DAY, the most mentioned word was “data.”
As mentioned earlier, Huan Mo believes that the industry is entering a new era with data-driven as its core.
And this new era is referred to by Huan Mo CEO Gu Weihao as ” Autonomous Driving 3.0.”
In the era of Autonomous Driving 1.0, the number of hardware determined the level of ability.
Many companies achieved autonomous driving by stacking hardware such as LiDAR, but this approach not only has only mediocre results but also adds to the cost of the vehicle, making it difficult to achieve mass production and landing. The range of autonomous driving test mileage is only about one million kilometers.
With the emergence of AI technology and the appearance of large-scale computing central chips, autonomous driving has also entered the 2.0 era.This is the stage where the majority of companies are currently at. Besides significantly improving driving performance and reducing vehicle costs, autonomous driving mileage has rapidly increased to millions of miles. However, meeting the demand for technological development is still difficult.
The autonomous driving 3.0 era is approaching technological maturity, with data beginning to enable self-training and allowing autonomous driving mileage to soar to over 100 million kilometers. Currently, only a few car companies such as Tesla can achieve this result.
In China, Momenta can be said to be the closest company to the autonomous driving 3.0 era.
The challenge lies not only in how to mass-produce and sell millions of cars like Tesla, but also in collecting and mining vast amounts of data brought by the 3.0 era.
To address this challenge, at the AI DAY in December last year, Momenta released the first data intelligent system in China called MANA.
In terms of architecture, MANA consists of four major components: TARS, LUCAS, VENUS, and BASE.
BASE is the bottom layer of the entire system architecture, including data foundations, data fusion, PoseidonOS, and more.
The other three components are located at the upper level:
-
TARS represents the development prototype algorithm of Momenta, including perception, planning decision, map positioning, and simulation engine.
-
LUCAS represents the perception, computation, verification, and other processes of vehicles in reality, that is, the large-scale generalization of autonomous driving.
-
VENUS is the data dashboard used to evaluate the quality of the algorithms based on a reference standard.
As mentioned earlier, the data mining is automatically completed by LUCAS, which coincides with Tesla’s efficient data processing. The reason why it can be achieved is that it has set up a set of intelligent data models similar to MANA.
As Momenta’s intelligent driving enters the urban scene, challenges such as frequent urban road maintenance, large and dense vehicles, narrow lane changes, diverse urban environments, and more are not far away.
Gu Weihao summarizes that this presents six challenges at the technology level:
-
How to apply large models in the autonomous driving field.
-
How to make data more valuable.
-
How to use heavy sensing technology to solve real space problems.
-
How to use human-world interaction interfaces.
-
How to make simulation more realistic.
-
How to make the movement of the autonomous driving system more human-like.At this HAOMO AI DAY, MANA has undergone updates and upgrades in perception intelligence and cognitive intelligence.
Firstly, MANA has greatly improved its model training effectiveness by using self-supervised learning methods with large-scale unannotated data from mass-produced cars.
Secondly, MANA has enhanced its perception capabilities and built an incremental learning training platform, which saves a lot of computing power.
Furthermore, by using a time-series transformer model to make virtual real-time map-building in the BEV space, MANA has achieved more accurate and stable output of lane marking perception.
Moreover, in the simulation system, MANA has introduced high-value real traffic flow scenarios, which are more timely and realistic, effectively solving the “difficult problem” of passing through city intersections.
It is worth noting that in addition to perception, MANA’s cognitive intelligence has also ushered in a new stage. By deeply understanding the massive human driving covered across the country, learning common sense and gesture personification, MANA makes semi-assisted driving decisions more like human driving behavior, and can choose the optimal route based on actual conditions to ensure safety and give the feeling of being with an experienced driver.
At the same time, the new powerful capabilities introduced by MANA and the Transformer deep learning model have brought tremendous consumption of computing power.
Gu Weihao mentioned that “Attention ” is a new trend in AI development.
This is because of its concise structure, which can stack basic units infinitely, and obtain enormous parameter models. At present, it has reached the level of tens of billions, and even trillions.
With the increase of parameters and the improvement of training methods, the effect of the large model has also steadily improved, and it has surpassed the average level of humans in many NLP tasks.
However, at the same time, it has been found that with the introduction of the attention model into the field of autonomous driving, the demand for computing power far exceeds the Moore’s Law. This results in high training costs for large models, especially difficult to deploy on terminal devices.
“Generally, the required computing power for Transformer is 100 times that of CNN, but under this computing power, an average of 6.9% of computing power contributes 94% of the value, and there are many weakly correlated and low-value operations in multiplication and addition operations and power consumption, resulting in a lot of waste.”
To this end, MANA is reducing the cost of autonomous driving using low-carbon supercomputers, improving the design of vehicle-side models and chips to achieve the deployment of large models on the vehicle side, and enabling large models to exert greater effect through data organization.As a super player who also relies on data intelligence, Tesla’s solution to its computing power issue is the supercomputer Dojo, which was officially unveiled in 2021, taking autonomous driving to a new level.
Almost at the same time, Momenta announced in December of last year that it was building its own supercomputing center.
And today, the supercomputing center is officially unveiled.
Gu Weihao stated that the goal of Momenta’s supercomputing center is to meet the needs of billion-scale parameter models, train data sets with a scale of 1 million clips, and reduce overall training costs by 200 times.
“In addition, we have effectively organized and improved the data distribution based on Momenta’s scene library, and built an incremental learning engine based on the continuous stream of production vehicle driving data. Through the use of sparse activation and operator deep optimization technologies, we continue to optimize training costs,” Gu Weihao explained.
From large-scale models to data intelligence systems, to supercomputing centers, if Tesla is leading the industry’s development abroad, then Momenta is carrying the banner of autonomous driving in China. The top players on both sides of the Pacific have chosen the same path for key technology points, making people look forward to what big moves each will make next.
Winning the First Large-Scale Production of NOH in China, Momenta Takes the Next Step
This year’s competition in the field of autonomous driving, in urban road scenes, is the most lively.
Following XPeng’s first shot in the city with NGP, Momenta also launched NOH, becoming the second company in China to enter the urban scene.
Subsequently, startups such as Li Auto, NIO, Jidu, Avatar, and SAIC IM all followed suit and installed urban assistance driving functions on their latest models.
Technically speaking, among all of the players above, there are mainly two schools of thought: one is the “perception fusion + high-precision map” camp, the other is the “heavy perception, light map” camp. Momenta belongs to the latter, launching the first urban assistance driving solution focused on perception.
In Momenta’s view, although high-precision maps can provide rich prior information, they cannot keep up with the demand for accumulated data to feed intelligent driving in large-scale, wide-scale Chinese cities due to freshness and regulatory issues.
This embodies the first principle of Momenta’s technological strategic decision-making, namely that a technology path that can quickly translate data scale advantages into capability advantages is a good path.
For the perception part lacking high-precision maps, Momenta uses Transformer-based neural network models to perform front-end fusion in three dimensions: space, time, and sensing, to improve algorithm accuracy.Translate the following Chinese text written in Markdown into English Markdown text while preserving the HTML tags inside the Markdown, and only output the result.
It is worth noting that Transformer neural network technology was almost simultaneously introduced to the self-driving industry by Tesla and Haomo Zhixing, demonstrating the tacit agreement between the industry leaders in the two regions.
At the previous HAOMO AI DAY, Haomo showcased its creative “dual-stream” perception model, which achieved traffic-light recognition under light map conditions and lane recognition through multi-sensor fusion on urban roads using its self-developed BEV Transformer technology.
This time, Haomo has brought six new features to NOH:
1. Intelligent recognition of traffic lights, covering various Chinese signal lights of different shapes
After multiple scene simulations, Haomo’s NOH urban model can recognize traffic signals, enabling the car to “stop at red lights, drive at green lights, and slow down and pass through yellow lights” accordingly.
To make users familiarize themselves with the operation, they need to lightly step on the gas pedal while the light is green.
2. Intelligent left- and right-hand turns
The vehicle’s left- and right-hand routes are based on the human experience. The system is designed to avoid active pedestrians and non-motorized vehicles during turning.
3. Intelligent lane changing
The car can automatically change lanes based on navigation guidance and higher traffic efficiency. It can also assess the movement of traffic participants in the rear and the available space for lane changes to ensure safe and smooth lane changes.
4. Intelligent avoidance of static obstacles
The car can accurately identify obstacles such as cones and roadblocks and slow down or detour them. If there is sufficient space for detours, the car will steer around the obstacle. If not, it will slow down and wait for the right time to proceed.
5. Intelligent avoidance of dynamic obstacles## 6. Intelligent Traffic Flow Management
The NOH system in the Haomo City will provide the ability to smartly evade obstacles at high speeds. For dynamic obstacles such as moving vehicles, the system will reduce speed appropriately and then, based on the feasibility of diversion space, choose either to slow down and follow or to choose a diversion, thereby ensuring safety and efficiency.
This system will mimic human driving behavior and predict the intentions of the driver in front of it by monitoring turn signals and brake lights. As a result, the system will operate more closely to how a human drives, thereby improving passenger comfort.
Which type of vehicle will be the first to install the NOH system from Haomo City?
In fact, the answer was revealed in August at the Chengdu Auto Show with the launch in the all-new Mocha DHT-PHEV model with a laser radar version.
According to Great Wall Auto’s plan, the car will be mass-produced in September, go on sale later this year, and be available immediately. This will also mark the first official rollout of the world’s first mass-produced NOH automatic driving system.
Although not the first to disclose the NOH feature, Haomo’s advanced sensory perception technology has allowed it to land in China first with a “mass-produced NOH” system.
This is perhaps the best testament to the “Haomo model.”
Meanwhile, Haomo’s end-of-line logistics automated delivery vehicles have also broken through the cost barriers.
Haomo Intelligent’s logistics automated delivery vehicle, “Little Camel 2.0,” has a single bike price of 128,800 yuan, making it China’s first mass-produced unmanned delivery vehicle priced at over 100,000 yuan and ready for deployment in urban and park environments.
According to Zhang Kai, the “Haomo Little Camel 2.0” has seven core functions, including L4-level autonomous driving, remote driving, low-cost deployment, vehicle management platform, remote monitoring platform, order management platform, and WeChat mini-program. The vehicle can efficiently execute order delivery, ranking it among the industry’s leading companies, and is expected to produce 10,000 vehicles a year.
After 1000 days, Haomo arrives at a new historic turning point, where Zhang Kai likened it to “from stumbling to aspiring.”
The past 1000 days have been a microcosm of the speed of development in the automated driving industry. Over the past decade, automated driving technology has seen accelerating evolution.
Now, the era of automated driving 3.0 is here. Both old and new participants in automated driving are trying to grasp certainty from this transformational technology.
Clearly, after going through the test of smart driving mass production and delivery, and completing the continuous evolution of AI automated driving technology driven by data, Haomo Intelligent, as seen from its successful ticket to 3.0 era after Tesla, is a predictable player.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.