Whether it is delivery speed or delivery quality, for Millima Zhixing, 2021 is the year when it ranks first in the mass production capacity of Chinese autonomous driving companies at the fastest speed.
This year, Mo Mo not only launched "China's first data intelligence system" - MANA (Snow Lake), but also took the lead in the industry to propose the "Mo Mo Win Formula" for automatic driving landing:
(Stable mass production capability * leading data intelligence * security) ^ ecology, while transforming technology into engineering, explore the formation of a unique autonomous driving commercial landing paradigm.
Entering 2022, Milli mo will continue to lead the pace of mass production and help China's autonomous driving continue to develop and leapfrog:
In terms of data intelligence, the data intelligence system MANA learning time is close to 200,000 hours, and the virtual driving age is equivalent to 20,000 years of driving time of human drivers;
In terms of passenger cars, two generations of HPilot products have been launched at the end of the year, and in March the Tank 500 equipped with the HPilot 2.0 driver assistance system was officially launched. By April 2022, the mileage of assisted drivers exceeded 7 million kilometers, and the total use time was 130,000 hours.
In terms of unmanned distribution, the baoding terminal logistics automatic distribution workshop was expanded to 10,000 square meters, which can achieve the production capacity target of 10,000 terminal logistics automatic distribution vehicles with an annual output of 10,000 units, more than 1,000 off-line vehicles, and more than 40,000 orders for Wumei Multi-point Supermarket in Shunyi, Beijing.
On April 19, the 2022 HAOMO AI DAY arrived as scheduled, announcing the rapid progress of the latest strategies, technologies and products, and the evolution and upgrading of MANA once again attracted much attention.
With a unique model and speed, It has set off the heat of China's autonomous driving year, showing the ambition and forward-looking planning of The End.
Behind the leading industry is a full understanding of the first principle of automatic driving, with MANA data intelligence to drive the evolution of automatic driving, leading millimeter into the era of full competition for automatic driving, and Tesla's automatic driving mass production showdown will also officially kick off.
01
"Dare to stand, dare to stand, dare to act"
Three predictions and three battles at the end of 2022
In December last year, at the first HAOMO AI DAY, Zhang Kai, chairman of Zhixing, made ten predictions for the 2022 autonomous driving industry:
It is mentioned that 2022 will be the most critical year for the development of the autonomous driving industry, the competition in the field of passenger car assisted driving will officially enter the second half, and the autonomous driving of other scenarios will also officially enter the first year of commercialization.
At this HAOMO AI DAY, Zhang Kai continued his previous prediction and once again made three major trend judgments for 2022:
In 2022, more details are expected to be issued at the national level to further regulate the ownership and safety of autonomous driving data;
Urban NOH further broadens the application scenarios of intelligent driving systems and pushes the experience of intelligent driving to a new height;
The automatic distribution of terminal logistics is on the eve of the outbreak, and the head customers begin to deploy on a large scale of the scene.
Since the beginning of the year, the end of the year has also launched three key and significant battles in 2022: the battle of data intelligence technology, the battle of assisted driving city scenes, and the battle of the scale of terminal logistics automatic delivery vehicles.
The three major predictions and three battles of 2022 are not only the general trend of the intelligent driving track, but also the industry insights that continue to be deeply cultivated.
So what are the goals and progress?
The first is the battle of data intelligence technology.
With the acceleration of intelligent driving from system prototype construction to large-scale mass production, the basic technology of the industry is undergoing key changes - Transformer big model conquest image vision, successively obtained Tesla and the terminal platform.
The rapid evolution of Camera has brought about a skyrocketing data scale, and the data will directly determine the mass production capacity. The electronic and electrical architecture technology has evolved from distributed to centralized, and the large-capacity vehicle specification chip has opened the first year of the car.
Data intelligence has also become the ideological seal of Zhixing since its inception.
MANA was born here, and rapidly grew into China's first autonomous driving data intelligence system, has become the core driving force for continuous evolution, the large-scale mass production products and massive data accumulation continue to transform into productivity, further enhance the ability of perception, cognition, labeling, simulation, calculation and other five aspects, thereby reducing development costs, improving iteration speed, escorting the other two battles.
The second is the battle of the assisted driving city scene.
Assisted driving technology is constantly updated, the scene is expanding, and the more frequent, more important and more complex urban environment is the next key scenario that intelligent driving must face after breaking through the high-speed scene.
With the release of the urban intelligent pilot assisted driving system of the head enterprise, this year is destined to open the second half of the competition in the urban scene.
In the United States, Tesla FSD shows the all-round performance of data volume, algorithm and architecture at a limited cost, while many parties believe that this technology route may face national security, localization and reliability constraints in China in the future.
In the urban environment, NOH will realize typical urban functions such as automatic lane change overtaking, traffic light recognition and control, complex intersection traffic, unprotected left and right turn, etc. according to the navigation route, and cope with complex urban traffic scenarios such as vehicle close entry, vehicle blockage, intersection, roundabout, tunnel, overpass, etc., with a 70% intersection pass rate and a 90% lane change success rate to achieve Chinese localization performance that is not lost to Tesla.
Finally, there is the battle for the scale of the terminal logistics automatic delivery vehicle.
The terminal logistics distribution scenario ushered in explosive opportunities under the epidemic, and with the labor cost problem after the disappearance of the future demographic dividend, the growth of the track is also optimistic in the future in the medium and long term.
Zhixing has now fully upgraded its terminal logistics automatic distribution vehicle production base, the new production line is designed according to the concept of "flexibility + customization", and equipped with MES production management system to achieve production transparency, management mobility, and decision-making data.
The second generation of terminal logistics automatic delivery vehicle Little Devil Camel 2.0 was also officially released this time - priced at 128,888 yuan, becoming China's first 100,000 yuan terminal logistics automatic delivery vehicle product.
Little Magic Camel 2.0 is equipped with a vehicle-level perception kit, ICU3.0 large computing platform, and a 600L large cargo space cargo box:
Based on MANA empowerment, it can achieve full coverage of complex traffic scenarios such as mixed traffic and congestion, and support the medium and low speed full road conditions of urban open roads;
Support rapid power exchange, to achieve 60 km of real endurance;
Provide intelligent voice, touch and other multi-mode interaction.
There is no doubt that Little Devil Camel 2.0 will further accelerate the scale and commercialization of terminal logistics automatic delivery vehicles.
02
WHERE:
"Perceptual Intelligence", "Cognitive Intelligence" and "Cost and Speed" have been comprehensively upgraded
With "data intelligence" as the core, the three major businesses of passenger car assisted driving products, low-speed unmanned vehicle ecological platform, and intelligent hardware are constantly rotating as three blades, continuously collecting scene data, and realizing positive development based on data intelligence.
At this HAOMO AI DAY, the most important release is also the continued evolution of MANA in terms of "perceptual intelligence", "cognitive intelligence", and "cost and speed", so as to calmly cope with the upcoming urban scenes in the future.
Gu Weihao, CEO of Mimo Zhixing, said on AI Day that the evolution of MANA's perception intelligence is mainly reflected in typical applications of urban scenes such as traffic lights and lane line recognition.
For example, intersection traffic light recognition is one of the most complex problems in cities - in AI applications, it is a typical small object detection, not only the state will change dynamically, but also the form varies greatly from place to place, and the most important thing is the difficulty of tying the road. The vehicle perception system must recognize the traffic lights from a distance in a very short time, correctly perceive and react accordingly.
To this end, through image synthesis and transfer learning, the technology iteration is accelerated, and the synthetic data is used to the maximum extent (that is, the difference in probability distribution between the massive synthetic data and the real data is reduced), so that the objective function f has the smallest prediction error in the real scene, and the call rate of traffic light recognition is further improved.
Millome also created a unique "double-stream" perception model, which decomposes the traffic light detection and the road binding problem into two channels, and uses a spatial attention mechanism to combine the two, thereby outputting the traffic light traffic status of the target lane after the road is tied, so that the daily passenger car test of millume realizes the traffic light recognition under the light map.
Lane line recognition is another major problem of urban intelligent driving, compared with the clearly marked highway lane line, urban road traffic is more dense, driving, parking on the lane line recognition has a great impact, some of the city lane line also has serious wear, complex identification, irregular edge shape and other problems.
Large models, represented by Transformer, come in handy here.
Last year, Tesla and Millima Zhixing gradually applied Transformer-based perception algorithms to lane line detection problems almost at the same time.
Through the self-developed BEV Transfomer, multi-sensor fusion lane line recognition is carried out on urban roads, showing the advantages of self-driving attitude tolerance, complex road surface longitudinal error, road surface ups and downs robustness, detection field of vision and other aspects compared with traditional models, and is expected to bring exponential efficiency improvement to the landing of visual algorithms on various intelligent driving product lines.
Coincidentally, Tesla also used Transformer in its latest FSD update.
The significant change in this update is also the upgrade of lane geometry modeling from dense rasters (a set of points) to autoregressive decoders, allowing the system to directly predict and connect "vector space" lanes point by point using Transformer neural networks.
The cognitive intelligence evolution of MANA is reflected in the application of micro-model upgrade and Ali M6 big model.
Gu Weihao said that taking the common left-turn scene waiting and observation as an example, the traditional handwriting rules and parameters are replaced by machine learning models, which not only avoids the failure of logic explosions and perturbation factors when there are more and more rules, but also deals with more complex small scenes and makes decisions more generalizable.
MANA also greatly improves the interpretability and generalization capabilities by applying the Alibaba M6 model. The scale of M6 large model parameters has reached 10 trillion, and it was previously mainly used in natural language understanding, text automatic generation and other fields.
The cooperation with Ali is the first encounter between the autonomous driving field and M6.
The cost and speed of MANA has also doubled.
Labeling massive datasets is another big challenge beyond the size of the data and the barriers are higher. Manual labeling by data labeling companies or platforms not only requires a lot of manpower and money, but also the process is very slow.
To be low-cost and efficient, automatic screening and annotation of data is critical.
It followed Tesla to build a closed-loop automatic annotation system, and used unsupervised automatic annotation algorithms to greatly improve the efficiency of data annotation.
Based on the automated data processing capability of the MANA system, the automation rate of the AI at the end of the minute has reached 80%, and in terms of implementation method, it is carried out through two stages: "target rough positioning" and "fine estimation of attributes", which greatly improves the labeling efficiency and reduces the cost of labeling.
At the same time, the PAI-EFLOPS team of Momo Zhixing platform and alibaba cloud PAI-EFLOPS team cooperated to implement distributed training of Swin Transformer models based on 128 card A100 clusters, which reduced model training costs by 62%, accelerated by more than 96%, and throughput exceeded 40,000 samples per second through large model training optimization.
In addition, it is worth mentioning that MANA also adds privacy protection and data security guarantees to the existing processing network to fully protect data security.
The continuous evolution of MANA and the accelerated listing of noH in the millima city have formed a positive cycle.
Based on the small magic box 360T computing power, 144M cache, 200K + DMIPS CPU computing power and the whole car two lidar, 12 cameras, 5 millimeter wave radar formed by the whole vehicle perception redundancy, equipped with HPilot 3.0 of the city NOH is about to come out.
Gu Weihao also posted a NOH report card on AI Day: the intersection pass rate exceeded 70%, the success rate of lane change exceeded 90%, and the traffic flow processing capacity reached level 4.
Three firsts were displayed in a high-profile manner: China's first large-scale mass-produced urban driver assistance product, the first heavy perception of urban NOH, and the first most practical and efficient urban driver assistance product.
Different from Tesla's pure visual route, Miller will use vision and lidar for fusion perception to provide a basis for decision-making, and achieve a wider range of cities under the SD map - Milliper's double hundred plan will also be launched on a large scale with this technical solution, and there will be more than 100 cities in the next two years, carrying more than 1 million passenger cars.
03
The intelligent driving competition is about to enter the second half,
How to accelerate the construction of industry barriers?
With a number of leading companies successively announcing the urban intelligent driving landing schedule and releasing test videos, the intelligent driving competition has officially announced that it has entered the second half, and the Matthew effect is accelerating and highlighting under the background of the convergence of overall technical route choices.
At AI DAY, the core rules of the competition for intelligent driving in the second half were revealed:
Under the MANA system, the first thing is to effectively obtain high-quality data and create openly with customers.
If the second half of the intelligent driving competition enters the urban scene as scheduled, data will become the focus of competition.
It is constantly emphasized that whoever can efficiently and cost-effectively mine the value of data can become the king of competition.
Different from Tesla and other car companies that are full-stack self-developed, the millimeter of the combination model of "car companies + autonomous driving technology companies" is different from that of each company, and the data comes from the "milli-terminal model" of open co-creation with partners and deep binding for common development.
Guided by the windmill strategy, combined with the MANA data intelligence system and 6P open cooperation, Zhixing continues to cooperate with customers instead of supply mode, and wins the trust of customers to achieve a win-win situation.
To this end, in the field of passenger cars, Miller has specially designed the 6P open principle to provide 6 different levels of cooperation from full-stack solutions to source code, so as to create a more open Millipede ecology:
Partners can choose to adopt a full-stack solution, can choose to cooperate at the data intelligence cloud service level, can choose software, hardware or module level cooperation, and even choose prototype code level customization.
Second, inject user thinking into the intelligent driving development and iteration process.
Especially for complex urban scenarios, higher usage frequencies must match more stable user perceptions. Otherwise, there is a high probability that there will be more potential security risks than in high-speed scenarios.
The idea is to embed problems and solutions into the earlier development process, proposing that intelligent driving should allow users to accept and trust faster, and adapt to user habits faster.
At the level of enabling users to accept and trust intelligent driving faster, Miller has created the most understandable open guidance system, the most reliable HMI display system and the intelligent voice two-way intelligent driving interaction system, so that users can obtain a comprehensive and objective understanding of the intelligent driving function of the vehicle.
At the level of system adaptation to user habits, Miller has built a user habit analysis system on the car side, developed user takeover analysis tools in the cloud, and continuously improved the data structure through real-time data analysis, thereby continuously improving the user experience.
The last is to reduce costs and increase efficiency, unify the action goals of all employees, and maximize the degree of software reuse and the efficiency of each link.
At the end of last year, it was mentioned in the company's internal letter that in 2022, the assisted driving system will land on 34 models of Great Wall Motors.
This means that more than 70 projects are super high delivery intensity, and if you want to lead the competition in the second half, you must also think about how to solve the problem.
The answer given this time is: the process development and standardized delivery capabilities of intelligent driving systems.
In terms of software reuse, the unified intelligent driving software architecture and algorithm interface, a set of software algorithms to support different manufacturers, different interface types of sensors; a set of application layer algorithm codes to support different chips and different operating systems; a set of application software combinations to distinguish different model configurations through different configuration words.
In terms of efficiency improvement, the application layer software and the basic software development process are decoupled through rapid prototyping, and the two-way parallel research and development is carried out; the simulation test verification process with the design and development process of the intelligent driving system is highly integrated, so as to improve the ability of the intelligent driving software to integrate successfully at one time;
In the system calibration link, using the calibration data of dozens of mass-produced models that have been accumulated, reinforcement learning is used to automatically recommend calibration parameters, and virtual calibration technology is used to save a lot of repetitive real vehicle calibration work, effectively saving 50% of the calibration time of real vehicles and saving calibration vehicles for customers.
The competitive landscape of autonomous driving is changing dramatically, and Gu Weihao, CEO of Zhixing, has ended the evolutionary road of MANA in the spatio-temporal dimension of reviewing the evolution of global autonomous driving development.
"In this intelligent revolution, Chinese researchers and companies are not only not late, but they are always at the forefront of autonomous driving with the United States."
From the Transformer big model to the heavy perception scheme, from the data intelligence system to the intelligent driving of urban scenes, this is not only a showdown of automatic driving mass production, but also the convergence of the industry giants' development consensus and action plans in the competition.
With a more open industrial ecology, a wider range of partners and a clearer mass production plan, the end of 2022 is worth looking forward to.
"The Heart of the Car · The connoisseur said" trailer
On Thursday, April 21, 20:00-21:00, Box Cars founder and president Zhang Shuguang will be a guest car heart · Experts said, talk about "travel market nuggets: where is the opportunity point of the B-end track of new energy vehicles?"