After the DriveGPT lands, the volume lidar and computing power will become meaningless?

The cold winter sweeping through the autonomous driving industry has not passed, and the problems such as long development cycle, high cost and difficult landing have not been completely solved. However, a large number of car companies are undergoing radical transformation of electrification and intelligence, and the research and development of autonomous driving is an unavoidable important part.

However, the route of attracting consumers by autonomous driving technology and stories has gradually become impassable. BYD Wang Chuanfu also stood up and bombarded autonomous driving, which made the chill of autonomous driving even stronger.

In fact, the problem of automatic driving landing is the same for everyone, and no one can avoid this problem. However, car companies and autonomous driving companies can open the book on "development cycle and cost", because in an unfavorable environment, whoever can persist and maintain it with lower cost and higher efficiency may usher in the spring bloom.

In the cold winter of autonomous driving, can AI really reduce costs? This week, at the 8th Millie AI DAY, Millie Zhixing released the first automatic driving algorithm model DriveGPT, which applied GPT model and technical logic, and officially named it "Snow Fox Hai Ruo" in Chinese. The first model of DriveGPT is the brand-new Mocha DHT-PHEV, which will be mass-produced. With DriveGPT, the cycle and cost of autonomous driving development will be greatly shortened and reduced.

At present, with the development of automatic driving technology, there are two main ways of automatic driving training: real road test and virtual simulation test.

The biggest feature of real road test is that it can match the real traffic environment and simulate all kinds of complicated situations in real life. At the same time, the real road test can also find some abnormal or unpredictable situations, and improve the adaptability and reliability of the automatic driving system. But the real road test needs a lot of time and money, and it also involves traffic safety, laws and regulations, personal injury and other issues, which brings certain risks and pressures to the test process. All these will lead to the increase of test cost.

The other way is virtual simulation test, which tests the autopilot system in a computer simulation environment and trains the model through virtual scenes. Virtual simulation test can avoid safety problems and cost pressure in real road test, and at the same time, it can generate a large amount of data quickly, improve test efficiency and data volume, and is supplemented by many AI artificial intelligence technologies.

However, the data and scenes in the virtual simulation test are artificially designed, which may not fully reflect the complexity and uncertainty of the real road. Therefore, the virtual simulation test sometimes needs a certain degree of real road test to verify its results.

DriveGPT adopts the same transformer model and RLHF human feedback learning ability as ChatGPT. By introducing real driving scenes and human driving takeover data, the cognitive decision-making model of autonomous driving can be continuously optimized, thus reducing the cost of autonomous driving development.

Because DriveGPT is trained in a virtual simulation environment, it can save the safety problems and cost pressure in real road test. DriveGPT can generate a large number of simulation data for training the model, which can well simulate the complexity and uncertainty of real roads, thus ensuring the robustness and reliability of the model. At the same time, training in a simulated environment can also greatly save time and cost.

15 days can complete the task of 1 year, and still "earn extra money"? Compared with the traditional real road test, DriveGPT can obtain a large amount of data quickly and efficiently. The training process of DriveGPT is fully automated, which is not affected by the test time, environment and other factors, and greatly improves the test efficiency and data volume. This not only saves training time, but also improves the accuracy and robustness of the model.

DriveGPT itself can distinguish driving scenes from non-driving scenes, and can understand the driving environment. It can also be used for scene recognition and marking tasks, such as marking lane lines, traffic participants, traffic lights, road signs and other detailed information. Moreover, the price of each image recognition optimization has dropped from 5 yuan to 0.5 yuan, and the cost has dropped by nearly 10 times. When GPT-4 of OpenAI came out, its map recognition ability attracted our great attention, which was similar to its bottom.

AI automatic drawing is gradually replacing manual work, and the efficiency and cost of manual marking cannot be avoided. The machine needs no rest, and it hardly dazzles. The cost of automatic marking is less than one tenth of that of manual work in the past, and it can meet the demand of manpower for one year in half a month.

DriveGPT uses a lot of simulation data to train the model, which can well reflect the complexity and uncertainty of real roads. DriveGPT can better handle natural language, images and other data types, and can learn by itself, thus improving the complexity and accuracy of the model.

In addition, in terms of product iteration, DriveGPT can provide fast and effective feedback for autonomous driving developers, help them debug and optimize the system faster, and further reduce the development cost of system iteration.

The model trained by DriveGPT can be transferred to the real road test for verification, which further improves the safety and reliability of the system. The model trained by DriveGPT can well reflect the complexity and uncertainty of the real road, so as to adapt to various situations more quickly in the real road test. DriveGPT can also make the system be in multiple parallel universes at the same time, that is to say, it can do all kinds of driving situations that may happen when encountering similar situations again in advance, and the ability to predict the trajectory of people and vehicles has also been greatly improved.

DriveGPT's capabilities are not limited to the field of autonomous driving. We can see that the partners are School of Computer and Information Technology of Beijing Jiaotong University, Volcano Engine, Huawei Cloud, JD.COM Science and Technology, Qualcomm, NavInfo, Intel, etc. Of course, many of them are supplier partners, but NavInfo may be another part of the empowerment of DriveGPT.

The official announcement of NavInfo said that access to the milli-end DriveGPT Hairuoxue Lake can achieve continuous two-way empowerment. With the help of DriveGPT Snow Fox Hai Ruo algorithm, the level of map making automation can be improved. DriveGPT has a strong understanding of images and can be applied to map drawing. DriveGPT can use AI model to identify objects, especially buildings. Specifically, it can train through a large number of map data and satellite image data, and then use these data to identify, classify and label building information. At the same time, because DriveGPT adopts AI technology, its recognition accuracy and efficiency are higher than the traditional manual mapping method.

In addition to building identification, DriveGPT can also be used for other surveying and mapping tasks, such as road marking, terrain analysis, map updating and so on. Especially with more and more vehicles equipped with DriveGPT, they can even try to generate high-precision maps with high freshness. Although the final construction is a system that focuses on perception and ignores maps, DriveGPT is also to deepen this goal, but high-precision maps can be handed over to suppliers to do other things, and it is not just autonomous driving that needs high-precision maps.

Summary: Compared with the traditional automatic driving training mode, DriveGPT can save the safety problems and cost pressure in the real road test, at the same time, it can obtain a large number of data efficiently and quickly, accurately reflect the complexity and uncertainty of the real road, provide fast and effective feedback, and can be transferred to the real road test for verification. The next assisted driving scheme may make the cost of hardware and software lower.

DriveGPT has great application prospects. Although we have only seen GPT-class autonomous driving at present, other car companies and suppliers will certainly not miss this opportunity. At the moment when AI technology explodes, it may be more advanced than DriveGPT. Automatic driving training methods are also on the way.

Moreover, with the full access of AI, it may no longer be a popular way to rely on heap hardware compared with the number of lidar, the number of pixels and cameras, and the ability of computing chips. No one wants to pay for the low capacity and high cost brought by heap hardware.

This article comes from Luca Auto, the author of Easy Car, and the copyright belongs to the author. Please contact the author if reproduced in any form. The content only represents the author's point of view and has nothing to do with the car reform.