The 8th HAOMO AI DAY: The commercialization of Little Magic Camel is accelerating again! Yizhuang starts operations and implements 9 major scenes​

April 11##, China's only focus on autonomous drivingAITechnology Feast——The Eighth SessionHAOMO## AI ##DAY##attack. This sessionHAOMO AI##DAY ends with “HAOMO SPEED,AI SPEED” is the theme, releases the first in the industry Autonomous driving generative large modelDriveGPT,Chinese name "Xuehu Hairuo", will reshape the road of automobile intelligent technology; ecological aspects, Hao Mo officially announced that it has obtained 3 designated OEM contract , Commercialization takes a leap forward;at the same timelaunchedChina's first re-sensory、The city that does not rely on high-precision mapsNOHwill be mass-produced soon , first landed in Beijing , Shanghai, Baoding and other cities, Ranked first in mass-produced autonomous driving in China... ​

（Zhang Kai, Chairman of HaoMo Zhixing）​

# #发毛releases#A series of new technologies, new products, and new ecology …… demonstrates the continued upward and new vigorous trend. It has been established for three and a half years, and it has always maintained for AI Technology#Enthusiasm, Use "Haimo Speed" leads the development direction of large-scale autonomous driving models, big data, and large computing power in China, and sprints towards the autonomous driving 3.0 era. ​

##（ Hao Mo ZhixingCEOGu Weihao releases a large generative model for autonomous drivingDriveGPT, Chinese name "雪湖·海如"）​

##HAOMO AI DAY scene,

Generative large-scale autonomous driving model created by Hao MoDriveGPT Xuehu·Hairuo’s grand release, ##Become a Big ModelTechnology Implementing a new paradigm of autonomous driving. It establishes RLHF (Human Feedback Reinforcement Learning) technology by introducing driving data to continuously optimize the autonomous driving cognitive decision-making model. At this stage, is mainly used for Solving the cognitive decision-making problem of autonomous driving, The ultimate goal is to achieve end-to-end autonomous driving. Haimo DriveGPT Xuehu·Hairuo has started open cooperation with a limited batch of ecological partners, Beijing Jiaotong University School of Computer and Information Technology, Qualcomm, Volcano Engine, Huawei Cloud, JD Technology, NavInfo, Weipai New Energy, Intel, etc.Universities and enterprises join. ​

(Hao Mo Zhixing Chairman Zhang Kai, CEOGu伟珏、COOhoujun、CIOZhen Longbao at the HAOMO AI DAY scene ) ​

In terms of products, Hai Mo continues to lead No. 1 in mass-produced autonomous driving in China. China's first large-scale mass-produced, sensor-focused city NOH## will be launched in Beijing, Baoding, Shanghai and other cities in the first batch , and start generalization testing, and will be implemented in an orderly manner to 100 by 2024 city. The first new Mocha equipped with HPilot3.0#DHT-PHEV is about to be relaunched Pound is on the market, this is also the first launch of HaimoDriveGPT Xuehu·Hairuo The landing model fully ensures the industry leadership of NOH in every city. ​ecological level,

Haimo Passenger cars6P has made a major breakthrough in open cooperation, and has cooperated with 3家OEM factory Sign the fixed-point contract , related projects are being delivered. This is an important step in the commercialization of Haimo, comprehensively guarantees the rapid development of Haimaosituation. ​

张凯, Chairman of HaoMo ZhixingJudgment: “2023年 Intelligent driving products have entered an explosive period across the board, The large model has been applied in the car end Satisfaction has become an important measure of product competitiveness. HaiMo’s continuously improving six closed-loop data-driven capabilities will further accelerate HaoMo’s entry into the autonomous driving 3.0 era and form a corresponding moat .”##​ A comprehensive breakthrough in the four major battles of Haomo,

won three Large OEM Fixed-point contractAn important step towards commercialization ​##“ 2023

will be the year of smart driving sprint and the year of big exams." Event site , Zhang Kai shared his thoughts on with the title "HAOMO SPEED, AI SPEED" 2023 Judgment of the macro trend of autonomous driving development in 2023, as well as the latest progress of the four major battles and commercialization development. ​

Zhang Kai believes that smart driving productsare Entering a period of rapid growth across the board , 2023 is a very critical year. First of all, urban navigation assisted driving products will be mass produced around in 2023 Get in the car and work hard,City navigation for main playersauxiliary drivingThe product has entered the competition of real user coverage and multi-city implementation. Secondly, the commercialization of lineend-logistics automatic distribution industry will It has become the focus of autonomous driving companies. In the field of passenger cars, , smart driving equipped with integrated parking and parking functions Products will usher in a wave of front-loading mass production;automatic terminal logistics at Distributionfield,Terminal logistics automatic distribution vehicles are used in supermarkets, express delivery, etc. The scene is ushering in an explosion,In 2023, sustainable commercialization closed loop will be realized in these scenes . ​

## (Zhang Kai’s judgment 2

023 This year is the year of the autonomous driving sprint and the year of the big test) ​In the context of the industry’s full development and the launch of more products, Zhang Kai introduced Haomo

2023## The latest results of the four major battles in #. ​

## (Hai Mo continues to lead China’s mass production of autonomous driving) One)

​

First is the "battle for the king of installed smart driving capacity". The first new Mocha DHT-PHEV equipped with HPilot3.0 is about to be launched, and the second model equipped with HPilot3.0 Weipai Blue Mountain will also be released this year. HaomoHPilot has been equipped with nearly 20 models. User-assisted driving mileage exceeded 40 million kilometers,HPilot2.0 assisted driving daily averagemileageThe usage rate reached 12.6% . Overseas layoutaspects,Vehicles equipped with HPilot have been shipped to the European Union, Israel and other regions and countries, and have been delivered to users one after another. Next, they will be sold in the Middle East, South Africa, Australia, etc.The market is being launched one after another; at the same time, Haimo##HPilot will soon mass-produce the Mexican version and the Russian version. ​

#In March, the High-tech Intelligent Automotive Research Institute conducted annual pre-installation mass production database and designated vehicle model database data for comprehensive evaluation, and through

################################################################################################################################################## High-end intelligent driving system mass production share leadership award. Third-party data proves that Weimo is the absolute leader in mass-produced autonomous driving in China. Haimo continues to lead China's number one autonomous driving industry ############. ############​#######################（#########DriveGPT Snow Lake ·Hai Ruo############has############completed training based on 40 million kilometers of driving data#########)### ######​#########

The second is "MANA Large Model Peak Battle", China’s first autonomous driving data intelligence systemMANAArchitecturehas ushered in a full line of upgrades. To April 2023##, MANA has been studied for more than 560,000 hours, which is equivalent to 68,000 years for a human driver. HaomoDriveGPT Xuehu·Hairuo, Has completed training based on 40 million kilometers of driving data, with a parameter scale of reaching 120 billion. ​

## Based on the principle of "first, scale first", we will accelerate the victory of the NOH Hundred City War) ​

The third is "Urban NOH Hundred Cities Battle", China's first mass-produced, re-aware urban NOH, has started generalization testing in cities such as Beijing, Baoding, Shanghai, etc., and will be mass-produced and launched on the road to It will be launched in 100 cities in an orderly manner in 2024. Based on the principle of “safety first, user first, scale first”, Haimou will accelerate the victory of the urban NOH hundred-city war. ​

## (Little Devil Camel has covered

九Big scene, start diversified operations)​

Finally, "End Logistics Automatic Distribution Business Battle" , Haimo terminal logistics automatic distribution vehicle##Little Magic Tuo has been used in supermarkets, smart communities, and campuses to fulfill contracts Distribution, catering retail, airport patrol, universityteaching##express delivery Nine major scenarios including the Pier, the Smart Park area, and the atmospheric environmental impact assessment have started operations ,Accelerate commercialization closed-loop capabilities. 2023year3month, Little Magic Camel 2.0 obtained the vehicle code of Beijing Yizhuang unmanned delivery vehicle, and started operation in Yizhuang. Haimo has also become After the upgrade of the "Beijing Intelligent Connected Vehicle Policy Pilot Zone Unmanned Delivery Test Specifications", unmanned delivery vehicles are allowed to be carried out on public roads in Beijing's high-level autonomous driving demonstration zone. First company tested. ​

“Technological leadership is the basis for survival, I encourage youallTechnology R&DStudentsInvest in technological innovation.” In his speech, Zhang Kai once again emphasized Haimo’s firm determination to invest in technology research and development. Up to now, Haimo has obtained 164 patent certificates, and papers from top international academic conferences have been included in 6 articles, the latest two were selected into CVPR, one of the three top conferences in the field of computer vision recognition, and IEEE TIV, the world's first professional journal for intelligent vehicles. Hao Mo has posted all papers in GitHub Open source,Shared with the industry. ​

（Hai Mohas been with 3 家OEM factorySign the fixed-point contract, related projects are under delivery)​

At the scene, Zhang Kai also announced to the outside world Important progress of Hao Mo 6P open cooperation , currently has cooperated with 3家OEM factorysigned Fixed-point contract , related projects are under delivery. “Hai Mo always thinks that, Autonomous driving is a cutting-edge industry that advances and retreats together and shares results.Only healthy ecological partners can support the rapid development of Haimo ." Zhang Kai said. ​

##(MANA’s six closed loops to build a leading data-driven system)​

In addition,Hai Mo has always insisted on the scene We will develop balanced user experience design, artificial intelligence technology, and technical engineering capabilities, and continue to improve user experience in a data-driven closed-loop manner. Zhang KaiIntroduction,In three months, I have been working in data-driven six Achieving multiple progress in a large closed-loop system: ​

In terms of user demand closed loop, Feimo continuously analyzes driving scene data to improve strategies and provides experience feedback on new functions; in terms of R&D efficiency closed loop, Feimo deeply integrates data-driven concepts into product demand definition, perception and recognition. Intelligent algorithm development and other product development processes have improved the overall development efficiency by by by 30%; in terms of closed-loop data accumulation, diagnosis has been deployed on the vehicle end. Service data scenario tags cover 92% of driving scenarios .
In terms of data value closed loop, there is no trace大 The model is continuing to mine the value of data to solve key problems; in terms of product self-improvement closed loop, has been realized in no time The speed of after-sales problem processing is ten times more efficient than traditional methods, and after-sales problems can be located in as fast as 10 minutes. In two yearsEffectively exploreProduct improvement points, the problem closure rate reached 76 %；#In terms of business engineering closed-loop, Haimo has further improved the links from collection backflow, annotation training, system calibration, simulation verification, etc. The entire product development process from the final OTA release stage is an engineering closed loop. ​

Hai Mo DriveGPTXuehu Hairuo releases , will be reshaped Automotive Intelligent Technology Road Line​

In the field of AI autonomous driving, Feimo is a pioneer in the development and application of large models. Onsite, Haomo智行CEOGu Weihao Releases a large generative model for autonomous drivingDriveGPT,中文名雪湖·## Hairuo. ​

##（DriveGPT Snow Lake ·海如）​

HaomoDriveGPT Xuehu·HairuoEstablished by introducing driving dataRLHF (Human Feedback Reinforcement Learning) technology continuously optimizes the autonomous driving cognitive decision-making model. Its ultimate goal is to achieve end-to-end autonomous driving. At this stage, it is mainly used to solve the cognitive decision-making problem of autonomous driving. In the future, we will continue to integrate the capabilities of multiple large models into DriveGPT. Currently, HaomoDriveGPT Xuehu·Hairuo has implemented the model architecture and The parameter scale has been upgraded to 120 billion. 40 million kilometers of production vehicle driving data are introduced in the pre-training stage, and 50,000 manually selected difficult scenes are introduced in the RLHF stage to take over Clips. ​

The underlying model of ##DriveGPT Xuehu·Hairuo adopts GPT (Generative Pre-trained Transformer) generative pre-trained large model , unlike ChatGPT which uses natural language for input and output, the input of DriveGPT is a text sequence after perceptual fusion, and the output is a text sequence of an autonomous driving scene, which is about tokenizing the autonomous driving scene to form "Drive Language" ultimately completes tasks such as decision-making and control of the vehicle, obstacle prediction, and output of the decision-making logic chain.

##（DriveGPT

雪湖·海如）##​##DriveGPT Xuehu·HairuoThe implementation process is to first train the initial model by introducing mass production driving data in the pre-training stage, then complete the training of the feedback model (Reward Model) by introducing driving takeover Clips data, and then through reinforcement In the learning method, the feedback model is used to continuously optimize and iterate the initial model to form a continuous optimization of the autonomous driving cognitive decision-making model. At the same time, DriveGPT

雪湖·海如 will also be based on the prompts at the input end and the decision samples of the Haimo CSS automatic driving scene library. Train the model and let the model learn the reasoning relationship, thereby splitting the complete driving strategy into the dynamic recognition process of the autonomous driving scene, and completing the generation of understandable and explainable reasoning logic chains. ​##（DriveGPT Xuehu·Hairuo

Released

, the first launch model of the new Mocha DHT-PHEV)​

At the scene, it was announced that DriveGPT Xuehu·HairuoThe first model is the new Mocha DHT- PHEV is about to be mass-produced and launched. Gu Weihao mentioned that DriveGPT雪湖·海如 can be gradually applied to urban NOH, shortcut recommendation, intelligent sparring and getting out of trouble in the scene. With the blessing of DriveGPT雪湖·海如, vehicle driving will be safer; the movements will be more humane and silky, and There is reasonable logic to tell the driver why the vehicle chose such a decision-making action. For ordinary users, vehicles will become more and more like experienced drivers. Users will have a stronger sense of trust in smart products and understand that vehicle behaviors are predictable and understandable. ​

##（DriveGPT雪湖·海如will reshape the automotive intelligent technology route)​

Currently, Haomo DriveGPT雪湖·海如 has been officially opened to the public. For the cooperation with the limited first batch of customers, Beijing Jiaotong University School of Computer and Information Technology, Qualcomm, Volcano Engine, Huawei Cloud, JD Technology, NavInfo, Weipai New Energy, Intel, etc. have already joined. Haimo DriveGPT’s opening up and services will promote autonomous driving practitioners and research institutions to quickly build basic capabilities. ​

##DriveGPT Xuehu·Hairuo is open to the industry)​

Haomo DriveGPT Xuehu·Hairuo will work with ecological partners to take the lead in exploring the four major application capabilities of , including intelligent driving, driving scene recognition, driving behavior verification, and difficult scene escape. Currently, Hai Mo is gradually establishing a set of 4D Clips-based systems in the process of using data DRIVINGThe scene recognition solution is extremely cost-effective. In the industry, to give correct annotation results, a picture costs about 5 yuan; if you use DriveGPT雪湖·海如 Scene recognition service , the price of a picture will drop to 0.5 yuan. The overall annotation cost of a single frame image is only equivalent to 1/10 of the industry. Next, Haimou will recognize the image frames and 4D Clips The service will gradually be opened to the industry, which will significantly reduce the cost of data use by the industry and improve data quality, thus accelerating the rapid development of autonomous driving technology. ​

##（

DriveGPT

雪湖·海如#scene recognition capability is open to the public）​

Gu WeihaoIntroduction, Haomo in2023 年1#China's largest autonomous driving industry Intelligent Computing CenterMANA OASIS（雪湖·OASIS) This time # have ##, further supports the computing power of DriveGPT Xuehu Hairuo. First of all, Hai Mo and the Volcano engine were newly builtThe "full set of large model training assurance framework" realizes the minute-level capture and recovery capabilities of abnormal tasks, and can ensure continuous training of kilocalorie tasksnumber months without any abnormal interruption, effectiveguaranteed了Large model training stability; Secondly, Haimo has developed an incremental learning technology# with real data return as the core, and has extended it# to large model training, building We built a large model continuous learning system, independently developed a task-level elastic scaling scheduler, scheduled resources at the minute level, and the cluster computing resource utilization rate reached 95%; Finally, MANA OASIS reduces costs and increases efficiency by improving data throughput to meet Transformer large model training Efficiency

,By introducing the Lego operator library provided by the Volcano engine to achieve operator fusion, the end-to-end throughput is increased by 84%. ​MANA architecture upgrade across the board

###Open empowerment, visual perception ability##################Compared with Tesla############### ###​############

Hai Mo#After more than a year of application iteration##, this AI DAY also ushered in a comprehensive upgrade, Officially open for empowerment. Gu Weihao introduced , MANA computing basic service has been specially optimized for large model training in terms of parameter scale, stability and efficiency. , and integrated into OASIS;Secondly,MANAPerception and cognition related large model capabilities are unified and integrated into DriveGPT雪湖·海如 ; Third, a data synthesis service using NeRF technology has been added to reduce the cost of acquiring Corner Case data; At the same time targets a variety of The problem of rapid delivery of chips and multiple models optimizes heterogeneous deployment tools and vehicle model adaptation tools. ​

## (China’s first autonomous driving data intelligence system MANA Comprehensive upgrade,Open empowerment）​

（MANA visual perception ability progress）​

In addition, MANA’s visual perception capabilities continue to improve. On the one hand, it can learn three-dimensional spatial structure and image texture at the same time, and its pure visual ranging accuracy exceeds that of ultrasonic radar, BEV The solution also has greater versatility and adaptability; on the other hand, it can achieve single-pass and multi-pass purely visual NeRF three-dimensional reconstruction, making the road scene more realistic and almost invisible to the naked eye. After scene reconstruction through NeRF, Corner Cases that are difficult to collect in real environments can be edited and synthesized. On the basis of the original global perspective modification and adding lighting/weather effects, the ability to synthesize virtual dynamic objects has been added. Various Hard Cases can be synthesized on the original set motion trajectories to simulate the complex urban traffic environment with lower cost. Cost testing improves the boundaries of urban NOH capabilities and better cope with the complex urban traffic environment.

（MANA visual perception ability progress）

It is worth mentioning that in the face of one of the most difficult visual tasks in the industry - monocular visual measurement, following the Tesla Later, Haimo also took the lead in China to verify whether it could use fisheye cameras instead of ultrasonic radars for ranging to meet parking requirements. Haimou introduced the visual BEV perception framework into the car-side fisheye camera, achieving a measurement accuracy of 30cm within a range of 15 meters and a visual accuracy of higher than 10cm within 2 meters. Using pure visual ranging to replace ultrasonic radar in parking scenarios will further reduce the overall smart driving cost.

( Pure visual ranging accuracy can use fisheye Camera replaces ultrasonic radar)

Gu Weihao said, "We have always mentioned that feedback data based on real user scenarios can allow us to better optimize products and make product progress faster. All technologies must be transformed into products that are useful to people to be most valuable. Now, Haimo’s products are providing more value to users. ”​

##Heavy guests from industry, academia and research in the field of autonomous driving help out, and the number of ecological partners has reached nearly百家​

This AI ##DAY once again gathered a super luxurious guest lineup in the field of autonomous driving. Dong Yang, Co-Chairman of China Automotive Chip Alliance and Vice Chairman of China Electric Vehicles Committee of 100, Zhu Xichan, professor at Tongji University and director of the Institute of Automotive Safety Technology, School of Vehicles and Transportation, Tsinghua University Professor Cao Dongpu and other industry experts and scholars shared the latest progress in the autonomous driving industry and technology; Weipai CTOZhang Botao,Tank Product Director Gu Yukun, Chief Scientist in the Artificial Intelligence Field of Huawei Cloud, Academician of the International Eurasian Academy of Sciences, IEEE/CAAI Fel low天奇、Ai Hezhi, Senior Director of Product Marketing at Qualcomm, NLai Junjie, Senior Director of Engineering and Solutions for VIDIA China, Huang Wuling, deputy director of the Cognitive Intelligence Laboratory of the Xiongan Innovation Institute of the Chinese Academy of Sciences and deputy director of the Xiongan New District Digital Transportation Laboratory and other ecological partners sharedIndustry Latest insights,and cooperation with Haimo Condition. In addition, at the end of this quarter, the second comprehensive scientific investigation and research on the Qinghai-Tibet Plateau by Yingchi Technology and the Chinese Society of Automation , Waiting for more partners, the number of Haimo ecological partners has reached nearly 100. ​

## (Haomo ecological partners attended HAOMO AI DAY) ​

Dong Yang said, “China’s automobile industry has experienced great changes, from the era of introduction to the era of innovation, from the era of electrification to to intelligence. The technology of intelligent connected cars is developing rapidly. In the past six months, my special feeling is that "L3 is about to come out" and everyone needs to pay attention. We hope that private innovative enterprises represented by Hao Mo Zhi Xing can serve as the main force and pioneer in the innovation and development of Chinese brand automobiles.” ​

Zhu Xichan said, “The safe operation of autonomous driving cannot be separated from the improvement and guidance of policies, and it cannot be separated from the careful exploration and implementation of relevant enterprises in the process of technology research and development. Practice. Hao Mo Zhixing regards security as a system engineering and underlying logic. Both attitude and practical results are worth sharing with the entire industry."​

Cao Dongpu said, "I hope that in the future we can work with Haomo to carry out more cooperation around technology research and development, industry standard formulation, talent training, etc., and jointly contribute to the development of autonomous driving in China."​

Gu Weihao mentioned at the end of his speech ,It has been nearly three and a half years since the establishment of Hao Mo. Over the past three years, many things have changed, but Shimo's firm investment in technology has remained unchanged. Haimo peopleAlways love technology, be prepared and sprint with all our strength. No matter how difficult it is, I will never give up. What you love is separated by mountains and seas, and mountains and seas can be leveled. Technology will definitely change the world! ”​

##In speech,Gu Weihao revealed the secret in his speech 雪湖·The origin of Hairuo’s name ：##"'海如 ' comes from "Zhuangzi Autumn Waters", contains two mythical characters He Bo and 北 Hairuo. He Bo asked 北Hai Ruo, what is the difference between big and small?北Hai Ruo teaches that you don’t feel big because of the heaven and earth, and don’t feel small because of the smallest detail. Haimao Snow Lake·Hai Ruo means wisdom and tolerance, embracing all rivers, and contributing to the development of the industry. ”​

##

Zhang Kai said," Spring is warm and flowers are blooming, AI is lighting up the future.Haimowill continue Move forward with determination.”​

The above is the detailed content of The 8th HAOMO AI DAY: The commercialization of Little Magic Camel is accelerating again! Yizhuang starts operations and implements 9 major scenes​. For more information, please follow other related articles on the PHP Chinese website!