Millimeter released the Chinese name "Xuehu Hairuo" of the automatic driving generation model DriveGPT.
DriveGPT Snow Fox Hai Ruo, a big model of autonomous driving, established RLHF (Human Feedback Reinforcement Learning) technology by introducing driving data, and continuously optimized the cognitive decision-making model of autonomous driving. At present, it is mainly used to solve the cognitive decision-making problem of autonomous driving, and the ultimate goal is to achieve end-to-end autonomous driving.
Millimeter DriveGPT Snow Fox Hai Ruo has started open cooperation with the limited first batch of ecological partners, and universities and enterprises such as School of Computer and Information Technology of Beijing Jiaotong University, Qualcomm, Volcano Engine, Huawei Cloud, JD.COM Science and Technology, Siweituxin, Weipai New Energy and Intel have joined in one after another.
In terms of products, NOH, the first mass-produced and re-perceived city in China, will land in Beijing, Baoding, Shanghai and other cities for promotion testing. By 2024, it will land in 100 cities in an orderly manner. The first brand-new Mocha DHT-PHEV equipped with HPilot3.0 will be launched soon, which is also the first landing model of Millimeter DriveGPT Snow Fox Hai Ruo, fully ensuring the industry leading position of Millimeter City NOH.
On the ecological level, a major breakthrough has been made in the open cooperation of 6mm passenger cars. Fixed-point contracts have been signed with three OEMs, and related projects are being delivered. This is an important leap in the commercialization of nano-powder, which fully guarantees the rapid development trend of nano-powder.
Zhang Kai, Chairman of Millimeter Star, judged: "In 2023, smart driving products will enter a full-line explosion period, and large-scale models will be applied at the car end, and the frequency of use and satisfaction of car owners will become an important measure of product competitiveness. The six closed-loop capabilities driven by the continuous improvement of data will further accelerate the pace of millimeter entering the era of autonomous driving 3.0 and form a corresponding moat. "
Zhang Kai believes that smart driving products are entering a period of all-round rapid growth, and 2023 is a very crucial year. First of all, in 2023, urban navigation-assisted driving products will be mainly mass-produced, and the urban navigation-assisted driving products of major players will enter the competition of real user coverage and multi-city landing. Secondly, the integration of navigation and parking and the commercialization of terminal logistics automatic distribution industry will become the focus of autonomous driving companies. In the field of passenger cars, smart driving products equipped with integrated parking and parking functions will usher in a wave of pre-installed mass production; In the field of terminal logistics automatic distribution, terminal logistics automatic distribution vehicles have exploded in supermarkets, express delivery and other scenes, and in 2023, a closed loop of sustainable commercialization will be realized in these scenes.
The first mocha DHT-PHEV equipped with HPilot3.0 will be launched soon, and the second model equipped with HPilot3.0 will also be released this year. HPilot has equipped nearly 20 models. The mileage of user-assisted driving exceeded 40 million kilometers, and the daily average mileage utilization rate of HPilot2.0 assisted driving reached 12.6%. In terms of overseas layout, vehicles equipped with millimeter HPilot have been shipped to the European Union, Israel and other regions and countries, and will be delivered to users one after another, and will be put into markets such as the Middle East, South Africa and Australia. At the same time, Hmong HPilot will be mass-produced in Mexico and Russia.
In March, High-Tech Intelligent Automobile Research Institute conducted a comprehensive evaluation every year based on the data of pre-installed mass production database and designated garage. Through the data research of nearly 20 vehicles before milli, milli was awarded the annual leading prize of mass production share of high-end intelligent driving system. Third-party data proves that Millie is the absolute leader of mass-produced autonomous driving in China.
Following "Battle of the Peak of MANA Big Model", the MANA architecture of China's first automatic driving data intelligent system ushered in a full-line upgrade. By April 2023, MANA will study more than 560,000 hours, equivalent to 68,000 years of human drivers. Millimeter DriveGPT Snow Fox Hai Ruo has completed the training based on 40 million kilometers of driving data, with a parameter scale of 654.38+02 billion.
The third is the "Battle of NOH Hundred Cities". China's first NOH city which can be mass-produced and re-perceived, has been popularized and tested in Beijing, Baoding, Shanghai and other cities, and will be mass-produced soon. By 2024, it will land in 100 cities in an orderly manner. Finally, the principle of "safety first, users first, scale first" will accelerate the victory of NOH Hundred Cities War.
Finally, the "terminal logistics automatic distribution business war". The terminal logistics automatic delivery vehicle "Little Magic Camel" has been put into operation in nine scenarios, such as Shang Chao performance, smart community, campus distribution, catering retail, airport patrol, university education, express delivery connection, smart park and atmospheric environmental assessment, thus accelerating the closed-loop ability of commercialization. In March 2023, Little Magic Camel 2.0 obtained the vehicle code of Beijing Yizhuang unmanned delivery vehicle and started Yizhuang operation. Millimeter has also become the first company to allow unmanned delivery vehicles to test on the open road of Beijing high-level self-driving demonstration zone after the upgrade of "Beijing Intelligent Networked Vehicle Policy Pioneer Zone Unmanned Delivery Test Specification".
"Technology leadership is the foundation of survival, and all technology research and development students are encouraged to invest in technological innovation." In his speech, Zhang Kai once again emphasized Mimo's firm determination to invest in technology research and development. Up to now, Millie has obtained 164 patent certificates and 6 papers from top international academic conferences. The latest two papers 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 of smart cars. Millicent has opened all the papers in GitHub and shared them with the industry.
At the scene, Zhang Kai also announced to the outside world the important progress in the open cooperation of MM 6P. At present, fixed-point contracts have been signed with three main engine plants, and related projects are being delivered. "I always believe that autonomous driving is a frontier industry where * * * advances and retreats together and * * * enjoys exclusive results. Only a healthy ecological partner can support the rapid development of millimeter. " Zhang Kai said.
In addition, Millie has always adhered to the balanced development of scenario-based user experience design, artificial intelligence technology and technical engineering capabilities, and continuously improved the user experience in a data-driven closed-loop manner. Zhang Kai introduced that in the past three months, we have made multiple progress in the six closed-loop systems driven by data.
In the closed-loop aspect of user demand, continuously analyze the driving scene data and improve the strategy, and give feedback on the new function experience; In the closed loop of R&D efficiency, the concept of data-driven is deeply applied to product development processes such as product requirement definition, perception and cognitive algorithm development, and the overall development efficiency is improved by 30%; In the closed-loop aspect of data accumulation, the diagnostic service data scene tag is deployed at the vehicle end, covering 92% of driving scenes.
In the closed loop of data value, the big model continues to dig data value and solve key problems; In the closed loop of product self-improvement, the speed of dealing with after-sales problems is ten times higher than that of traditional methods, and the after-sales problems can be located in 10 minutes at the earliest. In two years, the product promotion points have been effectively excavated, and the closed-loop rate of the problem has reached 76%; In closed-loop business engineering, we further improved the closed-loop product development process from collection and recycling, label training, system calibration, simulation verification to final OTA release.
Millimeter DriveGPT Snow Fox Hai Ruo established RLHF (Human Feedback Reinforcement Learning) technology by introducing driving data, and continuously optimized the cognitive decision-making model of autonomous driving. Its ultimate goal is to achieve end-to-end autonomous driving. At present, it is mainly used to solve the cognitive decision-making problem of autonomous driving, and will continue to integrate the capabilities of many large models into DriveGPT in the future. At present, the model architecture and parameter scale of millimeter DriveGPT Snow Fox Hai Ruo have been upgraded, and the parameter scale has reached 654.38+02 billion. In the pre-training stage, the driving data of 40 million kilometers of production vehicles were introduced, and in the RLHF stage, 50,000 manually selected difficult scenes were introduced to take over the editing.
The bottom model of DriveGPT Hairuoxue Lake adopts GPT (Generative Pre-trained Transformer) model, which is different from the input and output of ChatGPT using natural language. The input of DriveGPT is the text sequence after perceptual fusion, and the output is the text sequence of automatic driving scene, that is, the automatic driving scene is symbolized to form a "driving language", and finally the tasks of vehicle decision adjustment, obstacle prediction, decision logic chain output and so on are completed.
The realization process of DriveGPT Snow Fox Hai Ruo is as follows: firstly, in the pre-training stage, the initial model is trained by introducing mass production driving data, and then the feedback model is trained by introducing driving takeover fragment data, and then the iterative initial model is continuously optimized by using the feedback model through reinforcement learning, thus forming the continuous optimization of the automatic driving cognitive decision-making model. At the same time, Hai Ruo, a learning tiger of DriveGPT, will also train the model according to the tips from the input terminal and the decision samples from CSS automatic driving scene library, so that the model can learn the reasoning relationship, thus splitting the complete driving strategy into the dynamic identification process of the automatic driving scene, and completing the generation of an understandable and interpretable reasoning logic chain.
At the scene, Millie announced that the first model of DriveGPT Snow Fox in Hai Ruo is the new Mocha DHT-PHEV which will be mass-produced. Gu mentioned that DriveGPT Snow Lake can be gradually applied to urban NOH, quick recommendation, intelligent sparring and relief scenes. With the blessing of DriveGPT Snow Fox Hai Ruo, the vehicle will be safer to drive; The action is more humanized and silky, and there is reasonable logic to tell the driver why the vehicle chooses such a decision-making action. For ordinary users, vehicles are more and more like old drivers, and users will have a stronger trust in smart products and understand that the behavior of vehicles can be predicted and understood.
DriveGPT Snow Fox Hai Ruo will take the lead in exploring four application capabilities together with eco-partners, including intelligent driving, driving scene recognition, driving behavior verification and getting rid of difficult scenes. At present, in the process of using data, Millie has gradually established a driving scene recognition scheme based on 4Clips, which is very cost-effective. In the industry, to give the correct labeling results, a picture needs to be around 5 yuan; If you use the scene recognition service of DriveGPT Snow Fox Hai Ruo, the price of a picture will be reduced to 0.5 yuan. The overall labeling cost of a single frame image is only equivalent to110 of the industry. Next, MM will gradually open the image frame and 4D clip scene recognition service to the industry, which will greatly reduce the cost of using data in the industry and improve the data quality, thus accelerating the rapid development of autonomous driving technology.
Gu introduced that MANA OASIS, the largest intelligent computing center in China's autonomous driving industry, was released on June 5438+ 10, 2023, which upgraded three capabilities in terms of computing capacity optimization, further supporting the computing capacity of DriveGPT Snow Fox. First of all, Millie and Volcano Engine have newly established a "full set of large-scale model training support framework", which realizes the minute-level capture and recovery ability of abnormal tasks, and can ensure the continuous training of thousand-card tasks for several months without any abnormal interruption, effectively ensuring the stability of large-scale model training; Secondly, the incremental learning technology based on real data feedback is developed and extended to large-scale model training, and a large-scale model continuous learning system is constructed. Independent research and development of task-level flexible scheduler, scheduling resources in minutes, the utilization rate of cluster computing resources reached 95%. Finally, MANA OASIS can reduce costs and increase efficiency by improving data throughput, which can meet the training efficiency of the Transformers model. By introducing Lego operator library provided by Volcano Engine, operator fusion is realized, and the end-to-end throughput is improved by 84%.
MANA, China's first autopilot data intelligent system, has been fully upgraded and officially opened for empowerment after more than one year of application iteration. Gu Hao Wei introduced that MANA computing basic services are specially optimized for large-scale model training in terms of parameter scale, stability and efficiency, and integrated into OASIS; Secondly, MANA's perception and cognition-related large model ability are integrated into DriveGPT Snow Fox Hai Ruo; Third, increase the data synthesis service using NeRF technology to reduce the acquisition cost of corner case data; At the same time, heterogeneous deployment tools and vehicle adaptation tools are optimized for the rapid delivery of multi-chip and multi-vehicle models.
In addition, MANA's visual perception ability continues to improve. On the one hand, we can learn three-dimensional spatial structure and picture texture at the same time, and the ranging accuracy of pure vision exceeds that of ultrasonic radar. BEV scheme also has stronger universality and adaptability. On the other hand, single-pass and multi-pass pure visual NeRF 3D reconstruction can be realized, and the road scene is more realistic, and the difference can hardly be seen by naked eyes. After scene reconstruction by NeRF, corner cases that are difficult to collect in the real environment can be edited and synthesized. On the basis of modifying the original global perspective and increasing the lighting/weather effect, the synthesis ability of virtual dynamic objects is increased, which can synthesize all kinds of hard situations on the original set motion trajectory, simulate the complex urban traffic environment, improve the urban NOH capability boundary with lower test cost, and better cope with the complex urban traffic environment.
It is worth mentioning that in the face of one of the most difficult visual tasks in the industry-monocular vision measurement, after Tesla, Millicent is the first in China to verify whether the fisheye lens can replace the ultrasonic radar for ranging to meet the parking requirements. MM The visual BEV perception framework is introduced into the car-side fisheye camera, and the measurement accuracy is 30cm in the range of 15m, and the visual accuracy within 2m is higher than 10cm. Using pure vision ranging instead of ultrasonic radar in parking lot scene will further reduce the overall cost of intelligent driving.
Easy car super evaluation system is heavily launched! Professional, hardcore and comprehensive car reviews are gathered, and the annual car model list is freshly released! Search for "Super Review" on Easy Car App, waiting for you to see it!