In just three years, from its inception to mass production, Zhixing has become one of the fastest growing autonomous driving companies, what is the secret of its rapid development in the short term? Some people say that it is the endorsement of the Great Wall, some people say that it is the support of AI bulls, and today we talk about the technology of Zhixing.
At the beginning of this year, the HPilot 2.0 of Zhixing was just launched on the tank 500, and now HPilot 3.0 is also about to be released. Recently, Technical Director Pershing revealed the progress and innovation of HPilot 3.0 at the technical and data levels.
In terms of hardware, HPilot3.0 has 360TOPS hash rate, which is also the hardware with the strongest computing power in the autonomous driving industry. The vehicle is equipped with 12 cameras and 2 lidar, 5 millimeter wave radar, 12 ultrasonic radar.
You can run without a high-precision map city road
The HIGH-precision map is like an assistant to the stepping point in advance, so that you can once again embark on a familiar path. With the high-precision map has the direction of travel, however, the high-precision map also has its limitations, on the one hand, its production cost is very high, on the other hand, it needs to be used by the national review map, and there is no complete version of the urban road high-precision map within the scope of policy permission.
At the same time, the demand for intelligent driving has been menacing, and it is imperative to achieve more intelligent assisted driving without high-precision maps.
HPilot 3.0 aims to realize this idea, focusing on urban autonomous driving scenarios.
First at the perceptual level. HPilot 3.0 can detect traffic lights in advance, according to Pang Bo, the first step is identification, the second step is to tie the road.
Due to the change in the diversity of traffic lights, long-tail data is processed more, and like most AI companies, Zhixing uses a simulation system. But there will always be a gap between simulation and reality, so "blended transfer learning" is used to address and bridge the gap between the two.
Pershing introduced, "The learning model gets both simulation data and real data. After learning, the effect on the real data can be fully utilized by the simulation data. For example, add the distance of two dataset migrations on Loss, and learn to make the probability distributions of the two datasets converge in this way.
Prior to transfer learning, synthetic data and real data were distributed across their respective dimensions. In this case, the learning effect is not ideal. After hybrid transfer learning, synthetic data and real data are almost pulled together. As a result, after the virtual simulation data is added, its performance in the real world can automatically become better. In this way, it is possible to quickly converge on a variety of traffic lights and complete the identification of traffic lights according to different traffic light states. ”
Regarding the road binding, it is to solve the problem of complex road conditions and the identification ability under a variety of traffic lights.
"In the case of a road where it is impossible to determine which traffic light to follow, HPilot also learns through a model. When you enter a picture, there will be a branch to deal with the traffic light detection problem, and check out the traffic light on the image. There is another branch that learns a Characteristic Map through the attention mechanism, expressing the relationship between the traffic light and the road structure.
In general, seeing a road shape, feature map heat map has a high probability of confirming the corresponding traffic light positions and then binding them together. Through these two methods, one is detection, and the other is to learn the scene through the model and bind, you can find the topology information of the traffic light on the road. ”
Traffic light recognition is only one of the recognition capabilities of HPilot 3.0, which can also solve the various challenges of urban lane lines through the Transformer network, which makes vehicles improve the perception of traffic lights and lane lines in the absence of high-precision maps in urban scenes.
From perception to cognition, it is more difficult to make vehicles judge and make decisions like human drivers, especially in China's urban roads.
In the face of challenges such as roundabouts, crowded traffic convergence, and speed change, Mi mo zhixing made a TarsGo Model, learning the actions of human drivers in these scenes through model learning and three-dimensional learning.
TarsGo provides models of the car end that require a lot of human data to train. In this regard, based on the framework of Ali PAI, The Wisdom Ofe and Ali Damo Academy jointly used the M6 model to carry out data mining for automatic driving data. M6 can not only be used to collect these mass production data transmitted by cars and users' cars, but also can use the data of life that human society continues to generate every day, and the data of other industries and postmarks to iterate the ability of automatic driving. Pershing said that "the application of M6 in the field of autonomous driving marks the universalization of AI model capabilities."
Supported by exclusive MANA intelligent data system
Whether it is perception or cognition, behind the support is a large amount of data, behind the two modules is relying on mana data intelligence system iteration and evolution.
The MANA data intelligence system pioneered by Zhixing has surpassed the cost and speed.
It is revealed that in terms of data labeling, the use of self-established automatic labeling capabilities can effectively reduce the cost of more than 80% compared with manual labeling, and usually the cost of data labeling accounts for two-thirds of the total cost, so this is a lot of expenditure. Only when the cost of labeling is reduced, the operation of the entire data intelligence system will be healthier. In terms of training, the overall cost was reduced by 62%, and the acceleration ratio was increased to 96%.
Establishing the ability of data self-labeling requires a lot of effort in the early stage. Pershing told Che Yunmei that automatic driving itself is a data-intensive and has high requirements for the labeling quality of data, so data labeling is the embodiment of core competitiveness.
To this end, Zhixing has also set up a separate algorithm team and engineering team for automatic data annotation. "Labeling requirements are more stringent than perception, 2D pixels may be within 3 pixels can be framed, 3D may be a few centimeters of error, this box should be accurate enough." And automatic driving is not only to frame this object, the attribute requirements of this object are also very high, whether the object is blocked or truncated, whether it is on the ground or on the road, whether it is riding a bicycle or riding a motorcycle, whether it is a tricycle or an old man' music, these must have a clear semantic mark. So in 2D, 3D and attribute annotation, we have a lot of optimized workload. Pershing said.
Therefore, the MANA data system provides a strong convenience for the subsequent product development and landing of TheLm zhixing in terms of cost and speed, so that the urban NOH function can be landed this year.
Compared with the progress of the urban auxiliary driving function of the same industry, The Wisdom Walk has seized the leading position. After the ability to assist driving in the layout of highways has been basically realized, the main battlefield has shifted to the urban scene.
In Pershing's view, the reason why Zhixing can take the lead in landing is that it adopts a heavy perception scheme. While other peers are still in the R&D phase, they can be deployed in large-scale cities that users can use. If it is a heavy map scheme, the urban HD map will be limited. The noh city OFSOth has SOPs in June this year, and can achieve effective deployment in more than 100 cities across the country, so it has great advantages in scope.
The second is mounted on the Large Wall's large-scale models. The overall deployment range is large, the number of models is large, and we are based on more data, and the speed of continuous iteration is also one of the competitive advantages.
Che Yun summary
Autonomous driving has swept from high speed to the city, facing far more technical problems than the public has seen, and it is still a long way to continue to solve and popularize automatic driving. The "three major battles" mentioned by Zhang Kai, chairman of Zhixing, are not the battles of the entire autonomous driving industry, such as the battle of data intelligence technology, the battle of intelligent driving city scenarios, and the battle of the scale of automatic logistics delivery vehicles at the end of the world.