Produced by | Tiger sniff car group
Author | Wang Xiaoyu
Edit | attentive
The header image | IC photo
Founded 825 days ago, he stared at Tesla.
On April 19, Great Wall Motor's self-driving company, Millima Chi Heng, released the "NoH of Millima City" function equipped with HPilot3.0. Gu Weihao, co-founder and CEO of Momo Zhixing, said: "This will be the first large-scale mass production of urban assisted driving navigation system in China. ”
Gu Weihao also said at the scene, "Our product solution has greater adaptability and can be extended to serve users in larger cities, rather than just staying in the scope of several cities." At that time, our products will cover all scenarios such as highways, urban roads, and parking lots. In the urban NOH, our product strength leads Tesla's performance in China. ”
Gu Weihao, co-founder & CEO of Mimu Chi Heng
However, there is no chance to compete head-on with Tesla in the same market. At present, Tesla's strongest fully autonomous driving function FSD (Full Self-Drive) is only in large-scale internal testing in the United States, and Tesla in China only supports the enhanced version of the automatic driving assist function (EAP).
However, for the new challenges that Tesla FSD may launch in China in the future, Gu Weihao told Tiger Sniff: "We are confident of winning." When asked at what level the MM NOH is compared to "Wei Xiaoli", Gu Weihao's answer is: "We haven't looked at too many aspects, we think Tesla is doing relatively well." ”
Obviously, there is no "Wei Xiaoli" in the eyes of the end, and he wants to chase Tesla.
Go Tesla's way
12 years ago, the Great Wall began to play autonomous driving.
As early as 2010, Great Wall Motors conducted research on ADAS (assisted driving) systems internally. In 2015, the self-developed intelligent driving system was dynamically demonstrated at a technology festival inside the Great Wall, and in September of the same year, the first Haval H9 equipped with ADAS system was launched. In February 2017, Wei Jianjun, chairman of Great Wall Motors, announced for the first time a smart pilot system called "i-pilot" to the outside world.
This naming method is based on Tesla. As early as 2013, Elon Musk cited the term Autopolit, the autopilot on airplanes, to describe Tesla's self-driving technology. Later, Weilai also learned to take a name called NOP - NIO Polit. Of course, today's Great Wall Motors' products have been renamed with a new smart pilot name, called "NOH- Navigation on HPilot".
This H refers to the millimeter of the ultimate wisdom action.
The NOH of the Great Wall is similar to the NOP of Weilai and the NGP of Xiaopeng, and it is a foreign name. But from a functional point of view, everyone is based on high-level assisted driving based on navigation. Although Musk has strongly opposed the use of abbreviations to describe things, in order to create a uniqueness of technology, domestic manufacturers still popular use this abbreviation of the function description.
For the average consumer, understanding the technical details behind these names is bound to require a certain learning cost. So the point that can reflect these technical differences is that in which scenarios can high-level assisted driving be supported?
Last year, everyone was competing for the "high-speed" track, and the high-speed navigation assistance function developed based on the high-precision map was launched. Those that have achieved mass production deliveries currently include Tesla, Wei Xiaoli, and the Great Wall. However, with the acceleration of mass production of lidar and large computing chips this year, everyone has begun to pour into the "city" track, and they want to liberate drivers on urban roads.
The official expression given to the city NOH is:
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.
It seems like a few simple scenes, but it is really quite complicated to break through one by one.
At least, Huawei was overwhelmed. The first to blow urban assisted driving into the sky was the Polar Fox Alpha S Huawei Hi Edition. Last year, Huawei's high-end autopilot ADS system, which successfully completed the circle-breaking marketing with only a road test video, freely shuttled around the city and easily avoided takeaway brothers. However, the Hi version of the model, which was scheduled to start delivering in December last year, has not been heard from.
When asked who is stronger than The NoH and Huawei's ADS. Gu Weihao quipped: "Huawei has not yet mass production, we are all compared with mass production." When talking about Xiaopeng, he said, "It also depends on what level Xiaopeng finally mass-produced, and it should be said that we are OK." In the competition with the technical routes of these domestic manufacturers, Gu Weihao has always emphasized a statement: heavy perception.
Heavy perception, the corresponding other route is, heavy map. Both of the above manufacturers have adopted a heavy map route - based on the location of the traffic light in the high-definition map, pre-aiming on the perception, and then identifying. Because, with the assistance of high-precision maps, vehicles can find the traffic light information corresponding to the current roads and lanes very well. You can even get the traffic light countdown by V2X.
But the biggest problem lies in the freshness of high-precision maps. Because the city's road data is huge, this leads to a large collection volume and a long update cycle. You know, China's intercity highways and urban expressways and so on add up to 300,000 kilometers, but the country's urban roads have nearly 10 million kilometers.
"Now the high-precision map update, or use a more manual way, relying on the collection car on the road collection, once a week can have a good." An industry insider at a self-driving company told Tiger Sniff.
If you do not completely rely on high-precision map assistance, rely on bicycle intelligence to carry out high-level assisted driving in the city, the core problem that needs to be solved is how to make the car understand the traffic lights? Because there are four typical difficulties in traffic light recognition: one is the detection of small targets, the second is that the status will change in real time, the third is that the form varies greatly from place to place, and the fourth is the difficulty of tying the road (how to bind the traffic light to the corresponding road).
In the previous 11.2-kilometer-long noH experience in the milli-city, the vehicle's perception of traffic lights has not been wrong throughout the whole process. In particular, when multiple traffic lights of different forms appear in the field of view in front of the vehicle. The NOH can accurately find the traffic lights corresponding to the vehicle, display the information on the screen, and finally perform the correct action.
For example, as shown in the figure below, there are two traffic lights on a straight road, one vertical traffic light on the left and one horizontal traffic light in the distance on the right. For people, it is possible to make judgments without effort, but for cars, it is not simple.
Here, in fact, tesla has been using a technical route that Tesla has been taking - deep learning. Gu Weihao proposed, "We must solve the problem of traffic light recognition. First, a lot of data is needed to train learning. Accelerating the iteration of technology through image synthesis and transfer learning is a way to quickly acquire accumulated data. Among them, the problem of mixing real data and synthetic data is the main technical problem. ”
A "double-flow" perception model for traffic light detection and tying is designed, which decomposes the traffic light detection and tying problem into two channels. Pershing, technical director of Zhixing, explains: "When you enter a picture, there will be a branch that deals with traffic light detection and detects the traffic light on the image. There is another branch that learns a Characteristic Map through a mechanism of attention, expressing the relationship between this traffic light and the road structure where the vehicle is located. ”
Finally, the attention mechanism (Transformer) is used to combine the two and output the traffic state of the target lane after the road is tied.
Of course, not only to solve the problem of traffic light recognition, but also to solve the problem of lane line detection, obstacle detection, road traffic sign detection and so on through the perception algorithm based on the attention mechanism. In this way, as much visual perception as possible is used to complete the identification of road information and reduce the dependence on high-precision maps.
This is also why, at the press conference, I dare to set up a Flag that "there will be more than 100 cities where the city NOH function will land". It should be known that Xiaopeng's urban NGP will only be opened in some cities at the end of the second quarter. Similarly, Huawei's ADS is currently only conducting high-level autonomous driving function tests based on high-precision maps in Cities such as Shanghai and Shenzhen.
According to the product plan given by the Great Wall, it is expected that by the end of 2022, the Mimu Zhixing Auxiliary Driving System will land on 34 models of Great Wall Motors, accounting for about 80% of its overall models to be listed. Over the next two years, the total number of passenger cars equipped with the system will exceed 1 million units, making it the only company in China that can reach this scale in the short term.
But scale doesn't tell the whole story.
The siege of the "city" began
Technically, with the blessing of lidar and large computing chips, it is no longer difficult to use the automatic driving system to safely send people from point A to point B. What really determines whether you can bring software functionality to market is the user experience. Because it is not easy to make artificial intelligence drive in exactly the same way as human driving.
In the process of experiencing the NOH of the city, the biggest confusion lies in the brake feeling. Compared with humans, the brakes controlled by machines will have obvious frustration. When faced with traffic lights such as slowing down and stopping, usually we humans ourselves will loosen the throttle, slowly move forward by idling, and stop with the last foot. However, the machine brakes in advance, reduces the speed to a safe range, and finally brakes again.
The most typical scene in the city
"Now that I perceive this piece, everyone's accuracy rate is 99.999%. But in regulating this area, neither industry nor academia has found a methodology for criticizing good or bad. An industry insider in the autonomous driving industry told Tiger Sniff that the frustration of the brakes may come from the problem of regulation.
Zhen Longbao, chief delivery officer of Zhixing, told Tiger Sniff, "This is a normal performance. This is also the case when we test the car ourselves, and we welcome this kind of problem when we test the car ourselves. Because only every time we encounter a problem, it is our happiest time, far happier than a normal brake, because this means that we know the direction of the system correction, or know where the problem of the system is. ”
As for the optimization of the sense of experience, Zhen Longbao said, "Based on the rapid iteration of MANA (data intelligence system), including based on the current data closed loop, the user data feedback that has now been mass-produced, and the communication with users, we are rapidly iterating on the next generation of algorithms." ”
When the author previously experienced the Robotaxi of Wenyuan Zhixing and the Robobus of Light Boat Zhixing, there was no obvious brake frustration. However, considering that Wenyuan Zhixing and Qingzhou Zhixing are both calibrated to the L4 level of automatic driving capabilities. While relying on high-precision maps, the scene is constantly polished in the same range and on the same route, which has a better riding experience.
However, unlike the above-mentioned autonomous vehicles that operate in limited areas, functions such as the noH of the millima city, after being pushed to the user, the user will use it at random time and random place. You even need to consider different user statuses, you may be in a hurry to work today, you can't drive too slowly; today the weather is good, want to open slowly and so on. Obviously, this is no longer a purely technical problem.
This also reflects the trend of autonomous driving technology in the entire passenger car field - software research and development, from one delivery to a lifetime iteration. Then, through the ability of data to close the loop, continuously optimize the user experience, has become an eternal topic of this technology.
As a simple example, L2-level assisted driving, which is now highly popular, still has many flaws in the delicacy of vehicle control and the accurate grasp of user feelings. For example, many auxiliary drivers known as L2, the steering wheel will swing frequently from side to side, and it will not be able to walk in a straight line. Because in terms of lateral control, such as the centering overshoot, the number of convergences and other parameters, are currently within a broad standard, there is no mature technical standard to limit. Therefore, it is necessary to combine subjective feelings to narrow the technical standards.
This may also be one of the possible reasons why Xiaopeng City NGP announced the road test video early, but it was delayed in officially delivering to users. Also including Weilai, as early as January 2021, it launched the concept of NAD Nio's automatic driving, but so far it is believed that the official road test video will be released, not to mention open to users and media for a real road experience.
Li Bin has said before: "The development of NAD, a new generation of autopilot system NAD, is progressing smoothly, and we believe that NAD will fully surpass the experience of mainstream autonomous driving systems in the market." He also emphasizes the experience, not the cold data such as the model and mileage.
However, there are always two sides to everything.
With the existing technology and experience, it is difficult to land a person driving in the city that experiences a very good manned autonomous driving, but it is easy to do the automatic driving of the carrier. Because the most direct thing is that the vehicle does not need to consider the user experience at all, and does not have to worry too much about the division of accident liability. It also generates stable cash flow, which is almost a race track for self-driving companies.
Like in Shanghai at present, Meituan, JD.com and other platforms are launching a large number of unmanned delivery vehicles. For example, in the Jiading campus of Tongji University, there are already unmanned delivery vehicles that are delivering three meals a day to more than 10,000 students on campus; for example, in the Jiading Huayiyuan community, Jingdong unmanned vehicles assist volunteers to complete the transportation of materials.
At the same time as the release of the urban NOH function carried on the passenger car, The Miller also released an unmanned delivery vehicle - the Millima Little Devil Camel 2.0, L4 automatic driving capability, with rapid power change, 60-100 km cruising range, intelligent voice and touch multi-mode interaction and other functions.
"Our positioning is to help the global low-speed unmanned delivery vehicle ecosystem." Hou Jun, COO of Zhixing, told Tiger Sniff that the positioning of the helper means that it exceeds 50% in the market share. "Our dream is to really help end logistics, and even the efficient operation of the entire logistics industry, and make up for the needs brought about by the lack of capacity and labor."
Like the key to the success of the city NOH, the main focus of the unmanned delivery vehicle field is the scale capacity - at present, the Baoding factory can achieve the production capacity target of 10,000 terminal logistics automatic delivery vehicles, according to the official statement of the terminal logistics automatic delivery vehicle: this is the world's largest flexible manufacturing base for terminal logistics automatic delivery vehicles.
As a reference, Neolithic, which is currently known as the world's first large-scale road operation and the largest number of unmanned vehicles delivered and deployed, only released nearly 1,000 unmanned vehicle operation data when a new round of financing was completed in August last year. That is to say, if the production capacity of the unmanned vehicle at the end of the millimeter is completely digested, the title of the first unmanned vehicle scale will be taken down by the millimeter again.
Write at the end
Admittedly, in the competition for the head place on the self-driving track, Zhixing is going all out. But this is destined to be a protracted battle, the wind direction of technology is changing rapidly, and the needs of users are changing rapidly. On the road to entrepreneurship, it is still necessary to maintain a sense of crisis at all times.
In July 2020, Wei Jianjun, chairman of Great Wall Motors, questioned himself: "Can Great Wall Motors survive next year?"
In August of that year, when Wei Jianjun participated in an industry forum, he once again talked about the transformation crisis of the Great Wall: "As a traditional car company, we are also learning from new forces such as Weilai, Tesla, and Xiaopeng, and the trend of intelligence is irreversible, so in the past 3 years, Great Wall Motors has been rapidly changing." At that time, he was sitting next to He Xiaopeng, chairman of Xiaopeng Automobile, and Li Bin, founder of Weilai.
If they weren't present, perhaps the Great Wall would have learned the goal, and Tesla would have been left.