With the blessing of MANA OASIS, the computing power is extremely optimized and the training cost is reduced by 100 times.

On January 5, 2023, the seventh HAOMO AI DAY was held in Beijing. At the end of the year and the beginning of the year, China’s autonomous driving annual event is coming. At this AI DAY, Haimo shared the steady results of the three major battles in 2022, looked forward to the global autonomous driving development trend in 2023, and released the latest results of Haimo's technology and products.

(The 7th HAOMO AI DAY was held in Beijing)

Haomo Intelligent Computing Center "Snow Lake·Oasis" (MANA OASIS) was released , which is the largest intelligent computing center in China’s autonomous driving industry, reaching 67 billion floating-point operations per second. With the blessing of MANA OASIS, the five major MANA models of Feimo have been newly unveiled and upgraded, and the vehicle-side perception architecture has achieved cross-generation upgrades. Feimo's technology stack layout continues to maintain a complete and leading trend, especially in the aspects of perception and cognition, leading the industry and leading large models , large computing power, and big data development direction, sprinting into the era of autonomous driving 3.0.

(HaoMo Zhixing cooperates with Huoshan Engine: to build MANA OASIS, the largest intelligent computing center in China’s autonomous driving industry)

Product level, in the first half of 2024, HaoMo Urban NOH will be implemented in 100 cities. This is China's first urban navigation assisted driving that can be implemented on a large scale. It will maintain its leading position by focusing on perception, large-model technical routes and applications, and user closed-loop data construction in 2025. HPilot has fully entered the fully driverless era.

## Zhang Kai, chairman of Haimou Zhixing, said that in 2023, Haimou will continue to bring safe, reliable and practical intelligent driving products to the market quickly and on a large scale, winning over a large number of users. recognition and affirmation.

The top ten autonomous driving trend predictions for 2023 are released, and the four major battles of 2023 have officially begun. , but three years is only the beginning of a long entrepreneurial journey." With the theme of "Spring Returns in Time, AI Illuminates the Future", Zhang Kai reviewed the development results of Haomo in 2022 and announced the development plan for 2023.

(HaoMo Zhixing Chairman Zhang Kai gave a keynote speech)

Zhang Kai said that "the three major battles of HaoMo in 2022 have come to a solid conclusion." HaoMo will continue in 2022 Steadyly ranked first in mass-produced autonomous driving in China, the third-generation HPilot products are installed in nearly 20 models such as Wei, Tank, Euler, and Great Wall Cannon, and the mileage of assisted driving users has exceeded 25 million kilometers; Haimo City NOH is China's first An urban navigation assisted driving product that can be mass-produced, the software version is currently in the delivery state; the terminal logistics automatic delivery vehicle has initially completed the commercial closed loop, with more than 1,000 units delivered, and the number of Xiaomotuo delivery orders exceeded 130,000, and the commercialization process It is speeding up across the board; MANA has more than 420,000 hours of learning, and its virtual driving experience is equivalent to 55,000 years of human drivers. In addition, Feimo's self-developed AEB algorithm helped four models of Wei, Euler and Tank obtain five-star safety certifications from E-NCAP (European New Car Safety Assessment Association) and ANCAP (Australian New Car Safety Assessment Association). Equipped with Feimo HPilot 300 Mocha DHT-PHEV models have also been sent to Europe. Feimo is not only the first Chinese company to develop self-developed AEB algorithms at home and abroad, but also the first Chinese autonomous driving company to go overseas to Europe and Australia.

(HaiMo ranks first in mass-produced autonomous driving in China)

Behind the rapid iteration of products in the passenger car field, Haimo is created by Haimo The industry’s unique product capability iteration iron triangle: scenario-based user experience design, artificial intelligence technology, and technical engineering capabilities. Based on this, MANA has formed a powerful data intelligence-driven system, which includes six closed loops: user demand closed loop, R&D efficiency closed loop, product self-improvement closed loop, data accumulation closed loop, data value closed loop, and business engineering closed loop.

In the user demand closed loop, Haimo has established 5 major dimensions and 264 types of core user needs, and can extract core issues from a large number of core scenario data on a daily basis; the R&D efficiency closed loop uses simulation tools to improve R&D efficiency. At present, Haimou simulation covers more than 70% of R&D, and R&D efficiency has increased 8 times compared with two years ago; the product self-improvement closed loop has enabled Haimou's product self-improvement closed loop rate to exceed 70%, helping customers successfully achieve 7 OTA online upgrades; data In the closed loop of accumulation, Haimou will reduce data costs by 98% by the end of 2022 through self-developed automated data collection and automated data annotation tools; the closed loop of data value will allow Haimou to reach mass production in just 4 months by reusing and developing new models. , matching and calibration of new models are completed within 2 months, and the calibration efficiency ranks first in the industry; business engineering closed-loop allows Haimo to achieve a 100% one-time pass rate for intelligent driving products, making Haimo the only intelligent driving technology engineer in China The most experienced company.

(Haimo MANA’s six closed loops build a leading data-driven system)

Zhang Kai said at the scene that the six closed-loop capabilities of Haimo’s data intelligent drive system , assisting customers in the rapid large-scale mass production of smart cars, truly releasing value for users, and transitioning the smart driving system of smart cars from the early adopter stage to the user dependence stage. Haimo has become the first and only autonomous driving company in China to enter the rapid product iteration stage.

"In 2023, the competition in the second half of intelligent driving will enter a period of acceleration, and the commercial applications of high-end intelligent driving products will usher in large-scale implementation." Based on this basic judgment, Zhang Kai analyzed the industry, technology, products, Top ten predictions for the development of the autonomous driving industry in 2023 in terms of talent and other aspects.

At the industry level, autonomous driving will fully enter the 3.0 era in 2023; high-level intelligent driving will become standard equipment for mid-range models; a new generation of chips and sensors will enter the market, and the autonomous driving ecosystem will become richer and more diverse. At the technical level, the capabilities of large models in data synthesis, knowledge extraction, etc. will help autonomous driving iterate to achieve an order of magnitude improvement; competition in autonomous driving systems will shift from functional competition to commuting efficiency competition; supercomputing centers will become the entry-level configuration for autonomous driving companies . At the product level, urban navigation assisted driving has entered a stage of emphasis on perception, and large-scale mass production and delivery have begun; the user experience of intelligent driving will transform from early adopters to user dependence; the overall cost of terminal logistics automatic delivery vehicles has been reduced to less than 100,000 yuan, reducing Further improve production relations. In the field of talents, Zhang Kai believes that the fierce competition for talents in the field of AI autonomous driving will continue until 2023.

(Hai Mo releases the top ten trend predictions for the autonomous driving industry in 2023)

Faced with the trends and opportunities to be faced in the future, Zhang Kai officially announced that Hao Mo The 2023 development plan has not yet started the “four major battles in 2023”. The most important thing is the "battle for the king of intelligent driving installed capacity". Feimo will complete the asynchronous and parallel development of multiple platforms, dozens of models, and dozens of projects. Through the process development and standardized delivery of intelligent driving, we will further promote the future installation of Feimo assistance. The goal of driving the number of passenger cars in the product is one million; the second is the "Peak Battle of MANA Large Models". With the help of the intelligent computing center, Haomo will implement the application of large models to the end; the third battle is " "Urban NOH Hundred Cities Battle", Bomo Urban NOH will be implemented in 100 domestic cities in an orderly manner, leading the large-scale mass production of urban navigation assisted driving; and finally, "End Logistics Automatic Distribution Business Battle", Bomo will continue to adhere to 5S cooperation The service model helps partners quickly implement products and solutions and accelerate the application process of the unmanned delivery industry.

(Zhang Kai announced that Bo Mo has officially launched the “Four Major Battles of 2023”)

“Bi Mo always believes that great opportunities must be grasped by inner strength. , the key to winning in the fierce market competition is still to master the core technology." In his speech, Zhang Kai emphasized Haimo's firm determination to invest in technology research and development. Up to now, Feimo has obtained 146 patent certificates, covering fields such as data fusion, intelligent perception, intelligent decision-making, control execution, and model computing power improvement. "In the future, technology research and development will remain the top priority of Weimo's investment."

(Zhang Kai said that technology research and development will be the top priority of Weimo's future development)

MANA OASIS, China’s largest intelligent computing center for autonomous driving, was established, and the five major models were newly unveiled and upgraded

On the 7th HAOMO AI DAY, Haomo Zhixing announced the official establishment of the intelligent computing center "Snow Lake·Oasis" (MANA OASIS). MANA OASIS is the largest intelligent computing center in China's autonomous driving industry. It is jointly built by Haomo Intelligence and Huoshan Engine. It can reach 67 billion floating-point operations per second, a storage bandwidth of 2T per second, and a communication bandwidth of 800G per second.

(Gu Weihao said that MANA OASIS will help Haimo Sprint enter the autonomous driving 3.0 era)

Gu Weihao, CEO of Haimou Zhixing, said, "MANA OASIS Let Haimou have super computing power, abundant data and computing power, and Haimou's technical product capabilities will be even more powerful. With the support of 'MANA OASIS', Haimou will sprint into the era of autonomous driving 3.0."

President of Volcano Engine Tan Dai said, "We are honored to reach in-depth cooperation with Hao Mo Zhixing in the field of intelligent computing centers to jointly promote the leap-forward development of autonomous driving intelligent training platforms and accelerate the agile iteration and commercialization of autonomous driving technology."

(Hao Mo Zhixing Intelligent Computing Center "Snow Lake Oasis" reaches 67 billion floating point operations per second)

Data-driven is the direction and trend of autonomous driving development, massive Product data, ultra-large-scale data training, large model applications, etc. have all put forward greater demands for computing power. Based on the rich big data accumulation and underlying technology of the Volcano Engine, the computing, storage, and communication capabilities implemented by MANA OASIS allow data to be transformed into knowledge more quickly to achieve the purpose of increasing efficiency.

In terms of data management capabilities, in order to give full play to the value of the intelligent computing center and allow the GPU to continue to operate at saturation, after two years of research and development, Feimo established a full set of Data Engine for large-scale training, achieving a data screening speed of 100P Improved by 10 times, random read and write latency for tens of billions of small files is less than 500 microseconds. In terms of computing power optimization, Feimo cooperated with Huoshan Engine to deploy Lego high-performance operator library, ByteCCL communication optimization capabilities, and large model training framework. Software and hardware are integrated to optimize computing power to the extreme. In terms of training efficiency, based on Sparse MoE, through cross-machine sharing, large model training with 100 billion parameters can be easily completed, and the training cost of one million Clips (the smallest annotation unit of milliseconds of video) is only 100 kcal per week, and the training cost is reduced by 100 times.

(MANA OASIS basic capability display)

With the blessing of MANA OASIS, the five major models of MANA have been newly unveiled and upgraded to help the rapid development of autonomous driving: Among them, the visual self-supervision large model allows Haimo to be the first in China to realize automatic annotation of 4D clips; the 3D reconstruction large model helps Haimo to generate data, solve the data distribution problem at a lower cost, and improve the perception effect; multi-modal mutual supervision The large model can complete the recognition of general obstacles; the large dynamic environment model further uses heavy-sensing technology to reduce dependence on high-precision maps; the large human-driving self-supervision cognitive model makes Haomo's driving strategy more anthropomorphic, safer and smoother .

(MANA’s five major models help the development of autonomous driving capabilities)

First of all, the video self-supervision large model enables 100% automation of Haimo 4D Clip annotation. Manual labeling costs are reduced by 98%. In order to obtain more high-value data at a lower cost and more efficiently, it is necessary to solve the problem of automatic expansion from discrete frames to Clips form. Hao Mo first uses massive video clips to pre-train a large model through video self-supervision. It uses a small amount of manually annotated clip data to perform Finetune (fine-tuning) to train the detection and tracking model so that the model has the ability to automatically label; then, it will have The original video corresponding to the tens of millions of single-frame data labeled is extracted and organized into Clips, of which 10% are labeled frames and 90% are unlabeled frames. These Clips are then input into the model to complete the analysis of 90% of the unlabeled frames. Automatic annotation, thereby achieving 100% automatic conversion of all single-frame annotations to Clip annotations, while reducing the cost of Clip annotation by 98%. The generalization effect of the Haomo video self-supervised large model is excellent. Even in some very difficult scenes, such as severely obscured cyclists, small targets in the distance, bad weather and lighting, automatic annotation can be accurately completed.

(Visual self-supervised large model, can complete automatic annotation of 4D Clip)

Secondly, 3D reconstruction of large models allows Weimo to "create data out of nothing", and it is no longer difficult to obtain massive corner cases (long-tail scenes). Faced with the industry problem of "it is difficult and expensive to accumulate corner cases completely from real data", Feimo applies the popular 3D reconstruction NeRF technology to autonomous driving scene reconstruction and data generation. It changes the perspective, lighting, and texture materials. This method generates high-reality data, obtains normal cases at low cost, and generates various high-cost corner cases. The data generated by 3D reconstruction of large models is not only better and cheaper than the traditional method of manual explicit modeling and then rendering textures, but the addition of data generated by NeRF can also reduce the perceived error rate by more than 30%.

(3D reconstruction of large model, solving data distribution problem at lower cost, improving perception effect)

Third, multi-modal mutual supervision large model , allowing vehicles to grow "fiery eyes" and accurately identify special-shaped obstacles. After successfully achieving accurate detection of lane lines and common obstacles, Feimo is thinking and exploring more general solutions to the problem of stable detection of various special-shaped obstacles in the city. The large multi-modal mutual supervision model introduces lidar as a visual supervision signal and directly uses video data to reason about the universal structural expression of the scene. The detection of universal structures can well complement existing semantic obstacle detection and effectively improve the pass rate of autonomous driving systems in complex urban conditions.

(Multi-modal mutual supervision large model, can complete the recognition of universal obstacles)

Fourth, the dynamic environment large model can accurately predict roads The topological relationship allows the vehicle to always drive in the correct lane. Under the heavy-sensing technology route, Haimou faces the challenge of "real-time inference of road topology" in order to minimize its dependence on high-precision maps. To this end, based on the BEV (bird's eye view) feature map (feature map), Haimo uses the standard map as the guidance information, uses the autoregressive encoding and decoding network to decode the BEV features into a structured topological point sequence, and realizes lane Topology prediction allows Haimo's perceptual ability to realize real-time inference of road topology under the navigation prompts of standard maps like humans. Hao Mo believes that solving the intersection problem will actually solve the NOH problem in most cities. Currently in Baoding and Beijing, Hao Mo has an accuracy of 95% in inferring the topology of 85% of intersections. Even very complex and irregular intersections can be accurately predicted at any time.

(Large dynamic environment model allows Haimo to further use re-sensing technology and reduce dependence on high-precision maps)

Fifth, self-supervised recognition of human driving Know the big model, master the driving skills of high-level drivers, and make driving decisions smarter. In terms of exploring "using a large amount of human driving data to directly train the model to make anthropomorphic decisions", in order to enable the model to learn the excellent driving methods of high-level drivers, Feimo has newly introduced real user data, and at the same time used RLHF (from human The idea of ​​​​reinforcement learning in feedback first trains a reward model to select better driving decisions. In this way, the pass rate of Haimou in recognized difficult scenarios such as U-turns and roundabouts has been increased by more than 30%. This is the same idea as ChatGPT, which is popular in the AGI field, and uses human behavior feedback to select the best answer.

(A large self-supervised cognitive model for human driving, Hao Mo’s driving strategy is more anthropomorphic, safe and smooth)

MANA’s five major models comprehensively improve Hao Mo’s driving strategy Underlying technical capabilities systematized at the perceptual and cognitive levels. "With the help of five major models, MANA's latest vehicle-side perception architecture has integrated multiple downstream tasks that were scattered in the past to form a more end-to-end architecture, including general obstacle recognition, local road network, behavior prediction and other tasks. The car-side perception architecture has achieved cross-generation upgrades," said Gu Weihao. This also means that Wei Mo has stronger perception capabilities and stronger product capabilities, accelerating towards fully driverless driving.

(MANA’s five major models help Haimo sensing architecture achieve cross-generation upgrades)

In 2023, Haimo City NOH will be launched into large-scale mass production, and in 2025, it will be fully developed Entering the fully driverless era

At the scene, Gu Weihao officially announced the development plan of HPilot, BiMo’s assisted driving product: It is expected that by the first half of 2024, BiMo will complete HPilot’s plan to launch HPilot in 100 cities in China, achieving point-to-point interconnection. By 2025, NOH on a larger scale and in all scenarios will be implemented faster, ushering in a fully autonomous era.

(HPilot Development Plan)

Gu Weihao said that HPilot City NOH is China’s first urban navigation assisted driving that can be mass-produced. The leadership of Haimo City NOH is reflected in: leading the heavy-sensing technology route, leading technology application of large models, and leading user closed-loop data construction.

(Gu Weihao reveals the three leading advantages of Bomo NOH)

In the field of autonomous driving, important guests from industry, academia and research help to participate in AI DAY, Bomo is luxurious The circle of friends continues to expand

A luxurious lineup of heavyweight guests from industry, academia and research in the field of autonomous driving gathered at this HAOMO AI DAY. The Internet Society of China is the guiding unit of this HAOMO AI DAY. He Guili, Vice Chairman of the Internet Society of China, gave the opening speech; Wang Feiyue, Director of the State Key Laboratory of Complex System Management and Control of the Institute of Automation, Chinese Academy of Sciences, Yangtze River Scholar and National Outstanding Youth Fund Zhao Yao, a leading talent in scientific and technological innovation under the Ten Thousand People Plan and professor at Beijing Jiaotong University, shared cutting-edge insights into global autonomous driving technology; Tan Dai, President of Volcano Engine, Yang Jun, co-founder and CTO of Dada Group, Xu Tong, Product Director of Weipai Mocha, Ai Hezhi, senior director of product marketing at Qualcomm, Zeng Lei, product director of Euler Lightning Cat, and other partners shared the latest progress in cooperation with Feimo. Wu Xiangbin, director and chief researcher of the Intelligent Driving Laboratory of Intel China Research Institute, attended the speech.

(Haomo Zhixing partners attended HAOMO AI DAY)

Wang Feiyue said, “I hope that the Institute of Automation of the Chinese Academy of Sciences and Haomo Zhixing will work together to establish Haomo AI DAY. The technical system and theoretical system in the field of autonomous driving technology allow relevant scientific researchers around the world to share the results of Haimo."

Currently Haimao is working with the School of Computer and Information Technology of Beijing Jiaotong University on automatic annotation related technical fields. in-depth cooperation. Zhao Yao said, "I look forward to working with Haomo in the future to strengthen cooperation in talent training, technology research and development, and industry standard formulation, establish a technological moat, and work together with academia and enterprises to make China's autonomous driving bigger and stronger."

Yang Jun said that thanks to excellent domestic unmanned vehicle companies like Haimou, Dada and Haimou have joined hands to carry out very valuable unmanned delivery exploration and laid a good foundation for the development of the industry.

Wu Xiangbin expressed the hope that Hao Mo Zhixing would seize the opportunity, continue to innovate, be open and win-win, and contribute to intelligent connected cars with Chinese characteristics.

In addition, relevant government departments, professional institutions, and associations in Beijing, Shenzhen and other places, Alibaba, Meituan, Tencent, Baidu, JD.com, ByteDance, NavInfo, AutoNavi, Dixoncom, BMW, and Mercedes-Benz , Hyundai, BYD, Chery, Great Wall, Qualcomm, Intel, Renesas, Cambrian, Sagitar Juchuang, Haomo Investors and other partners participated in the live broadcast of this event online or offline.

Gu Weihao said at the end of his speech: "Every experience is a gift. The most winning weapon for Bimo is the entrepreneurial spirit formed by Bimo students to face difficulties together. This kind of invincible entrepreneurship Spirit is the biggest magic weapon for us to face challenges, take the lead, and keep moving forward. Wei Mo hopes to join hands with more fellow travelers to do more meaningful things together." At the end of his speech, Zhang Kai released a new vision of Wei Mo: "Let machines move intelligently, Give life more beauty." This is also the theme of the seventh HAOMO AI DAY. The confidence shown and the released content of this HAOMO AI DAY strives to light up the industry with the power of warmth in the cold winter. At the beginning of 2023, we will make full strides towards the new three years.

​​​

The above is the detailed content of With the blessing of MANA OASIS, the computing power is extremely optimized and the training cost is reduced by 100 times.. For more information, please follow other related articles on the PHP Chinese website!