Author| Xiangwei
Editor | Dexin
On April 11, Momo Zhixing officially released DriveGPT , a large generative model for autonomous driving.
The core of Xuehu Hairuo is to input various driving scenarios as tokens into the model, and then output a series of possible driving scenario tokens.
Translated into vernacular, it is to let the car recognize the road environment it is in and decide how to drive next.
The parameter volume of Xuehu Hairuo has reached 120 billion , which is close to the parameter scale of the GPT3 period.
Of course, learning a language is different from learning to drive a car, and there is no hard and fast standard for how many parameters can be used to train an autopilot to become a "veteran driver".
However, the emergence of Xuehu Hairuo means that domestic autopilot players have found a new entrance to improve their regulatory capabilities .
Dismantling "Snow Lake Hairuo", the number of parameters reached 120 billion
Before getting to know Xuehu Hairuo, let’s review ChatGPT first , the two are quite similar.
In 2017, Google first proposed a new learning framework - Transformer , which allows machines to learn a large amount of text at the same time. Compared with the serial learning of the previous RNN cyclic neural network, this new learning framework goes a step further and can achieve parallel learning.
In 2018, a company called OpenAI launched the Generative Pre-trained Transformer based on Transformer , referred to as GPT, the Chinese name is the generative pre-training model, which is the first generation of GPT- GPT 1 .
After investing more data and computing power, GPT 1 has experienced the evolution of GPT 2, GPT3, and GPT3.5.
Until November last year, it began to be able to talk to humans more intelligently, and under the promotion of Microsoft, an important investor in OpenAI, it became famous all over the world and became the super product that everyone is talking about today-ChatGPT.
ChatGPT is essentially a large language model. Input a word or text to ChatGPT, and the model will give the probability of the next word or text appearing. Finally, it can talk to humans because the model has learned a lot of human language, and can give a probabilistic result based on the language knowledge base and reasoning logic.
In the evolution of GPT, the size of the parameter scale is very important. The parameter volume of each generation of products has evolved from the initial 120 million to 1.5 billion and 175 billion.
During the evolution process, in order to keep the dialogue robot from talking nonsense, Open AI also added the artificial feedback mechanism RLHF, which can be understood as the engineer giving good or bad feedback to the robot to guide the dialogue robot to become smarter.
Today's ChatGPT can already talk to humans smoothly and naturally, and answer questions in various fields.
Can such a useful tool be used in the field of autonomous driving?
Haomo Xuehu Hairuo has become the first large-scale model product for autonomous driving in China. The difference from ChatGPT is that Hairuo is facing the driving language Drive Language .
The general operation process is to input a piece of environmental information of the previous N seconds into the large model of Xuehu Hairuo, such as the state of the vehicle itself, the state of surrounding obstacles, or the road environment, etc., and Xuehu Hairuo will generate the next environment that will happen What kind of results, such as the response measures of other road traffic participants and own vehicles, etc.
So how does the machine work?
The first step is conversion.
Autonomous driving sensors will collect various data and generate a plan result through the BEV perception architecture.
Xuehu·Hairuo will discretize the image perception results of BEVs in the entire space through the grid, and form a fixed-size vocabulary as a token by judging each grid. After this process, the lane line can be , road obstacles and other real-world information into a driving language that the machine can understand.
The second step is pre-training.
Regarding the pre-training model, Xuehu Hairuo is slightly different. It is adjusted from the encoder+decoder structure of ChatGPT to the GPT model of Decode-only structure, and is trained through 40 million kilometers of mass-produced car driving data .
In order to keep Xuehu Hairuo from driving indiscriminately, Momo also introduced humans to take over the data, and used about 50,000 Clips to train the feedback model and obtain a scoring model, so that the machine can finally better understand the environment and continue to learn Make better autonomous driving decisions.
It should be pointed out that the current Xuehu Hairuo is a large-scale model in the cloud, and the results of the large-scale model of Xuehu Hairuo will first land on the new .
If ChatGPT is a chat robot, it can interact through chat and give text answers that humans want. Xuehu Hairuo is more like a driving robot, allowing the machine to interact with the road environment more and give better driving decision-making answers.
Going forward along this path, Momo's automatic driving system will be more human-like, with a more mature understanding and processing of road scenes, and the end of the route may be the realization of end-to-end automatic driving.
Let the machine learn human driving, Xuehu Hairuo, a product based on the GPT large model, has closely tied the iteration of car intelligence with the evolution of artificial intelligence.
With Xuehu Hairuo, the cognitive ability of autonomous driving will be better improved, which in turn will help improve planning and control.
At present, Momo gives several major application :
Can be used to develop urban NOHs due to improved planning and control capabilities;
Can be used for shortcut recommendation;
Because it can deduce the path of some future road scene changes, Xuehu Hairuo can also act as a smart driver coach to help car owners drive better;
It can rule out some difficult scenes encountered in driving and carry out intelligent ambulance;
Xuehu Hairuo not only targets the autonomous driving industry, but also opens up corresponding capabilities and resources to industries such as robots, automobile OEMs, chip manufacturers, and scientific research institutions.
At present, Xuehu Hairuo has officially opened to the outside world and started cooperation with the limited first batch of customers. Intel and others have joined.
The first step of opening up is some data-related capabilities, and capabilities such as driving behavior verification and difficult scene escape will be gradually released in the future.
Mass production is progressing rapidly, breaking the circle and winning three car companies
In addition to the release of Xuehu Hairuo, the pace of mass production is also very fast. According to Zhang Kai, chairman of Momo Zhixing, the company will make progress in four aspects in 2023.
In addition to the release of Xuehu Hairuo, the pace of mass production is also very fast. According to Zhang Kai, chairman of Momo Zhixing, the company will make progress in four aspects in 2023.
Millisecond HPilot:
The first new Mocha DHT-PHEV equipped with Momo HPilot3.0 will be launched soon, and the second model Weipai Lanshan equipped with Momo HPilot3.0 will also be released this year.
At present, HPilot has been equipped with nearly 20 models as a whole. The user-assisted driving mileage exceeded 40 million kilometers, and the average daily mileage utilization rate of HPilot2.0 assisted driving reached 12.6%.
Overseas, vehicles equipped with HPilot have been delivered to users in the European Union, Israel and other regions and countries. It will also be launched in the Middle East, South Africa, Australia and other markets in the future. HPilot will also mass-produce Mexican and Russian versions.
MANA：
By April 2023, the learning time of MANA will exceed 560,000 hours, which is equivalent to 68,000 years for human drivers.
DriveGPT, the world's first generative large model for autonomous driving created by Momo, has completed training based on 40 million kilometers of driving data, with a parameter scale of 120 billion.
City NOH:
NOH in Haomo City has started generalization tests in Beijing, Baoding, Shanghai and other cities, and can carry out large-scale mass production and landing. In 2024, it will be launched in 100 cities in an orderly manner.
Zhang Kai believes that the mass production of NOH, which uses heavy perception and does not rely on high-precision maps, is more than a year faster than players in the industry.
Terminal logistics automatic distribution:
Xiaomotuo, the terminal logistics automatic delivery vehicle of Momo, has started operation in nine scenarios including supermarket fulfillment, smart community, campus delivery, catering retail, airport patrol, college education, express self-pickup, smart park, and atmospheric environmental impact assessment.
In March of this year, Little Magic Camel 2.0 obtained the vehicle code of the unmanned delivery vehicle in Yizhuang, Beijing, and started the operation in Yizhuang.
Momo has also become the first company allowed to test unmanned delivery vehicles on open roads in Beijing's high-level autonomous driving demonstration zone after the upgrade of the "Beijing Intelligent Connected Vehicle Policy Pioneer Zone Unmanned Delivery Test Specification".
At the scene, Zhang Kai also announced to the outside world the important progress of the 6P open cooperation. At present, fixed-point contracts have been signed with 3 OEMs , and related projects are being delivered.
According to HiEV's understanding, the three OEMs include OEMs other than Great Wall Motors.
Compared with its peers, Momo's technology iteration and mass production pace are relatively faster. It is the Tier 1 who knows the most about autonomous driving technology, and the one who knows how to mass-produce among the automatic companies. On the basis of accumulation, the players who have truly realized the data closed loop.
Looking back at the development process of Momo, you will find that its ambition is by no means to become a simple Tier1, but an artificial intelligence company.
The reasons behind it are closely related to the core team and technical architecture.
The executive team with Zhang Kai, Gu Weihao, Hou Jun, and Zhen Longbao as the core has integrated the genes of OEMs and technology companies. From the beginning of its establishment, the team has always adhered to the data-driven technical route.
Momo released China's first autonomous driving data intelligence system, MANA Snow Lake, and built an intelligent computing center, MANA OASIS Snow Lake·Oasis .
Based on the large-scale pre-installed mass production of the L2 assisted driving system, Momo has firmly established its position as the first in mass-produced autonomous driving, and has formed a data intelligent closed-loop system, realizing a virtuous circle in terms of iteration speed and cost optimization.
After more than a year of application iterations, MANA is now undergoing a comprehensive upgrade and has begun to empower the industry externally.
The entire technology evolution trajectory of Momo is essentially the process of an artificial intelligence company in the vertical field of automobiles, constantly releasing new technologies and quickly getting on the car.
If we compare Momo with Tesla, we will find that there is a large degree of similarity between the two, both of which are self-driven and constantly introducing new technologies.
As far as autonomous driving is concerned, from the BEV perception algorithm to the Occupancy Network, Tesla is gradually evolving through new technologies to allow vehicles to better perceive and understand the world and make better planning and control.
The only difference is that Momo’s technology is not only provided to itself like Tesla, and is more used for its own huge car fleet, but is industry-oriented, not only serving Great Wall Motors, but also continuously expanding beyond Great Wall Motors. Automobile customers.
Momo broke through the Great Wall system in the scope of cooperation with OEMs this time, which gave an important signal to the industry.
At present, the background of mass production in the field of smart driving is that, on the one hand, the development of Huawei's car BU has told us that there are many restrictions on providing smart driving solutions to car companies. On the other hand, some chip companies and smart driving companies are launching low-cost solutions such as integrated travel and parking, and have begun to receive cooperation orders from car companies.
These cases tell us that suppliers of smart driving solutions are facing difficulties in getting a car, and at the same time have the opportunity to cooperate with car companies.
"We did mass production three years ago, which is really hard work. Mass production is completely different from demo. Those companies that have not done mass production of pre-installation suddenly announce that they will enter the field of autonomous driving, and they will be in a difficult battle." Zhang Kai said with emotion after the press conference.
Therefore, it has mass production experience and can meet the needs of car companies for smart driving solutions. In this way, even a smart driving supplier with a background in a certain car company has a chance of survival.
Cracking the regulatory problem, Chinese players are accelerating their evolution
At a time when the voice of L4 autonomous driving companies is declining, OEMs have become staunch supporters of autonomous driving.
There are Teslas abroad, and domestically, there are important players in the field of autonomous driving such as Haomo, Wei Xiaoli, and Huawei.
Right now, these players are entering the deep waters of autonomous driving technology— planned control .
"The problem now is not that everyone doesn't know whether the method is good after making a method, but that they don't know how to do it." Ai Rui, vice president of technology of Momo Zhixing, said not long ago.
Similarly, when asked about making autonomous driving more like an old driver, which core module should be solved next, Wu Xinzhou, vice president of autonomous driving at Xiaopeng Motors, also put forward a similar view, thinking that the perception of Xiaopeng Motors is "until now The overall state is good, but the core and specific workload is still in forecasting and regulation.”
Wu Xinzhou himself is an in-depth user of Tesla FSD. After using FSD, he feels that CNGP is not inferior to the other party in terms of the delicateness of processing in many places. But what he is not sure about is whether the other party will be able to deal with the complex scene in China, especially in terms of regulation and control.
Yu Chengdong said more directly, "China's roads are very complicated. Tesla FSD is easy to handle in the United States and Europe, but it may be possible for them to do it in China."
Prediction and regulation test the cognitive ability of autonomous driving .
Veteran drivers with many years of driving experience will make a more comprehensive prediction of the road environment based on experience, and quickly make optimal driving operations.
The same is true for automatic driving, not only to perceive the road environment, but also to know how to drive like a veteran driver.
Tesla is at the forefront, and the FSD function has been applied in North American urban scenarios. Tesla's self-driving route has made a demonstration, leading car companies to advance to the throne of fully automatic driving.
Regardless of when fully automatic driving will come, or whether it will come, at least we have seen a consensus that car companies must invest in the research and development of automatic driving.
In particular, companies such as Haomo, Xiaopeng, and Huawei have invested in large-scale research and development of autonomous driving technology.
The new technologies released by various companies, such as the emergence of Xuehu Hairuo, not only let us see where Chinese autonomous driving players have gone, but also let us see the determination of car companies to invest.
ChatGPT, a generative dialogue product given by OpenAI, is a key to better interaction between human beings and the machine world. The same interaction between cars and the road environment in Xuehu Hairuo is a key to improving the regulation and control of autonomous driving. key.
Recently, it was reported that Tesla FSD will have a major update, and it may be on the agenda to introduce it to the domestic market in the future.
We are very curious about what kind of pattern Tesla and domestic autopilot players will present in China's road environment, whether they are in a state of following, or they will achieve technological catch-up in the regulation and control link.