This is not a six-month technological innovation, but once every 4 months, the end of the mouth said slow kung fu, the frequency of the foot can not be slow at all, the blink of an eye, the end of the release of the city NOH in China, from the point of time, faster than the new forces, even Tesla, and in the words of the end of the day, there is no new forces in the market, the old forces are divided, everyone is in a starting line.
Wen 丨 wisdom driving network Wang Shuoqi
It has only been 4 months since the last time the Mimi AI DAY, perhaps the high-speed NOH on the mocha has not been properly felt, and the noH of the city of the Mola has appeared in front of the eyes.
On April 19, Miller released the "NoH of Milliped City" equipped with HPilot3.0.
According to Milli, this is China's first large-scale mass production of urban assisted driving products, the first heavy perception of urban assisted driving solutions, and also the first most practical and efficient urban assisted driving products in China in 2022.
It is understood that this AI DAY is its fifth brand activity in more than two years, and before that, Zhixing once again completed the A+ round of financing of hundreds of millions of yuan invested by BOC and participated in by Shoucheng Holdings, also known as the national team entry.
And this speed seems to confirm the understanding of speed that millimeter has always had, slow is fast, and the young company seems to have been ahead of others in technical mass production.
First, let's look at a set of data, as of April, the actual total mileage of user driving assistance has exceeded 7 million kilometers, entering the first camp of data accumulation.
At the same time, at present, the assisted driving system has been installed to six models of Weipai Mocha, Tank 300 City Edition, Weipai Machiduo DHT, Weipai Latte DHT, Haval Divine Beast, and Tank 500, and the user's auxiliary driving mileage has exceeded 7 million kilometers. And in 2022, as many as 34 Great Wall models will be equipped with millima intelligent driving assistance functions, of which 30% are standard; and in the next two years, the number of Great Wall passenger cars equipped with millima intelligent driving assistance products will exceed 1 million units.
The second is the partner team of the terminal logistics automatic delivery vehicle business of TheLmot, represented by Ali Damo Academy, Wumart Multi-point, Meituan, etc., and under the epidemic in Shanghai, there are also many milli-terminal figures.
Today, the field of terminal logistics automatic delivery vehicles has also ushered in an update, the end of the little magic camel 2.0 officially released, the price of bicycles 128,800 yuan, becoming China's first 100,000 yuan level terminal logistics automatic delivery vehicles, and when this price was announced, there was also a burst of applause on the scene.
At the technical level, China's first data intelligence system MANA ushered in a number of heavy upgrades, perception capabilities by leaps and bounds, through the original "double-stream" perception model and self-developed BEVTransfomer, "so that China has no unrecognizable traffic lights and lane lines" has become a reality; cognitive ability, cost and evolution speed have also doubled, model training costs have been reduced by 60%, acceleration ratio of more than 96%, in addition to labeling AI automation rate has reached 80%, greatly reducing the cost of labeling costs.
Such an intelligent system, like a positive accelerator, continuous data is constantly optimizing the technology of the millimeter, better technology means more vehicle use, more vehicle use means that the accelerator runs faster.
The difficulty of the algorithm required for urban driving assistance is far greater than that of high-speed road conditions, which is why so far, there are still no mass production vehicles in China that can open navigation automatic driving assistance on urban roads.
——01——
How difficult is it?
First of all, we must know that urban NOH means that in the city, select the navigation, the vehicle can follow the planned route, and eventually reach the destination.
According to the prediction of Zhang Kai, chairman of Zhixing, in 2022, the competition in the automatic driving industry will officially enter the second half, mainly focusing on the pilot intelligent driving of the urban open scene: "In 2022, the state will issue more detailed rules to regulate the ownership and safety of autonomous driving data; the urban NOH will push the experience of intelligent driving to a new height; the automatic distribution of terminal logistics will be on the eve of the outbreak, and the head customers will begin to deploy the scene on a large scale." ”
It is claimed that the system can realize the main functions of automatic lane change overtaking, traffic light recognition and control, complex intersection traffic, unprotected left and right turning in the urban environment according to the driving route provided by the navigation, and can also cope with complex urban traffic scenarios such as vehicle close entry, vehicle obstruction, intersection, roundabout, tunnel, overpass and so on.
While this is an easy task for drivers, it is not for autonomous driving.
Take the traffic light, the domestic style is strange, the countdown is also different, and the end is mainly dependent on vision.
In interviews, how to tie traffic lights to roads is called "tying up roads" by engineers at the end of the road, and the autopilot system needs to recognize it and then match its signal to the road.
Millimae's choice is to use data to solve such problems, different traffic lights in various parts of the country are collected at a high cost, and millimeters are also combined with data simulation to speed up efficiency.
Through the simulation scene, a variety of different lighting, weather, angles, etc. are simulated to simulate the actual scene.
Through the data of these scenarios, it is fed back to the model so that the model can learn better.
Therefore, one after another different traffic lights, from identification to binding to the road, form a big data "heat map", these heat maps have become the textbook of machine learning.
This method is called the "double-flow" perception model by Millide, which decomposes the traffic light detection and the road binding problem into two channels, so that the daily passenger car test of Millimae realizes the traffic light recognition under heavy perception.
According to the calculation of the end of the millimeter, with the increase of the computing power of the car end, the large model represented by transformer will play an increasingly important role in the future automatic driving, and this advanced perception model is also used to identify the wrong and complex lane lines in the city.
Using Transformer's attention mechanism, it is possible to solve the stitching problem between multiple cameras very effectively, and 6 cameras all see the lane line, some next to it, some in front, and some in the back.
In addition, with the addition of attention mechanism, milli-mo can easily stitch together the perspective of multiple cameras to determine the position of itself and the lane, and then calculate that even if the lane line suddenly disappears, the milli-mo can cope with it freely.
Of course, all this is based on Mana, China's first data intelligence system.
——02——
MANA re-accelerates
According to the introduction of Miller, the MANA perception ability has made great progress, which is also highly correlated with the traffic light recognition ability and lane recognition ability in the city.
Behind the perception ability is the great changes in the perception technology of autonomous driving in the recent year, including the geometric growth of the chip's computing power, the emergence of the Transformer transmodal model and the rapid improvement of the Camera pixel, which are the basis for the development of technology.
In addition, there are also advances in "cognitive intelligence" and "cost and speed".
In terms of "cognitive intelligence", MANA replaces the traditional handwriting rules and parameters with machine learning models, which solves the problem of bloated code and easy to collapse and fail in the face of complex scenarios, makes scene decisions more generalizable and applicable, and greatly improves the interpretability and generalization capabilities.
It can be said that cognitive intelligence is mainly a BUG repair and efficiency improvement.
In terms of cost and speed, Miller and Alibaba have carried out in-depth cooperation in big model data processing technology, realizing the first encounter with the M6 model in the field of autonomous driving.
Finally, relying on the model cooperation of Ali Damo Academy, the model training cost was reduced by 60%, the acceleration ratio exceeded 96%, and the throughput exceeded 40,000 samples per second.
In addition, the automation rate of labeling AI has reached 80%, reducing a large number of manual labeling.
Such cooperation directly represents cost savings and efficiency increases.
Therefore, it can be summarized that the rapid growth of MANA, the rapid improvement of on-board hardware, its own perfection, and the help of external forces, but in any case, the end of the integration of advantages is gratifying from the perspective of efficiency.
——03——
Accelerated growth period at the end of the day
Urban driving assistance may be the most critical step on the eve of L3 level, in addition to the integration of urban and high-speed full scenes, more importantly, there is also the significance of commercialization.
Unmanned urban logistics can be opened, unmanned taxis can operate, flash delivery, takeaway, express delivery will no longer have people...
All of this requires urban driving assistance, and Millima seems to be leading half a position.
Tesla can now detect traffic lights, but it also does not control vehicles at intersections with traffic lights, and Wei Xiaoli actually does not have a full OTA for traffic light detection, which is the ultimate opportunity.
In addition, for the Great Wall, it may have won a final card at the commanding heights of automotive technology.
【Pay attention to new energy vehicles, pay attention to the video number of the electric laboratory】