Give life to time, give civilization to time
Mi mo Zhixing (the great wall's autonomous driving company) used the most circle-breaking sentence in "Three-Body" to end the fifth AI Day. As the CEO of Millima Chi Heng said, Mo Mo Zhi Xing is a "company that grows on the three-body".
Musk and Tesla have their "first principles," and Sodom Zhixing also has their "thought stamp." Each solution is based on a number of key assumptions, which can be called first principles or "thought stamps" or "thought stamps."
The "thought stamp" comes from Hines, the third wallfacer in The Three-Body Problem, and refers to a "key hypothesis" that can be implanted in the human brain to make it unconditionally agree.
Based on the "thought stamp" of identification, Zhixing has sorted out a set of intelligent driving solutions - data system MANA. Interestingly, the name MANA was also inspired by the three-body problem.
The fourth wallfacer, Luo Ji, walked on the frozen lake, fell into the ice cave, and on the verge of suffocation, he had an epiphany of the truth of cosmic sociology, summed up the laws of the dark forest, and transformed into a sword bearer, temporarily freeing the earth civilization from the threat of the trisolaran civilization.
This lake is the origin of the new era and has great significance in the three-body universe, but unfortunately Liu Cixin did not give this lake a name. There is a lake in Xinjiang, China, which is very similar to the three-body description, called Lake Manasarovar, and the data intelligence system is named after this lake.
Partners who are familiar with the three-body universe will surely marvel at the ingenuity of this set of naming.
After talking about the three-body universe of Millima Wisdom, let's review the latest progress released on this AI Day, the NoH assisted driving system in The Millima City.
As mentioned in the video, this is the third urban pilot assistance system opened to media experience after Huawei ADS and Xiaopeng XPILOT3.5 in China.
The system will be delivered through the new weY Mocha model in June this year, and will be installed in many Great Wall brand models such as WEY, tank, and Haval in the future. Of course, if other car companies are willing to use it, they can also cooperate to carry it.
Overtaking in curves that heavily sense the route
The biggest feature of noh in the city of Mimo is "heavy perception, not relying on high-precision maps". This is an innovative attempt in the field of intelligent driving.
High-precision maps can provide a wealth of prior information, which is bound to be more effective with half the effort, but at present, city-level high-precision maps not only have the problem of freshness that cannot be guaranteed, but also because of censorship problems that affect the rhythm of delivery.
Huawei ADS, Xiaopeng City NGP can only be delivered in first-tier cities such as Beijing and Shanghai, and the follow-up promotion speed is relatively slow, there are more than 300 cities in the country, and it is basically impossible to achieve city-level high-precision map coverage.
Because the scheme does not rely on high-precision maps, it is not interfered with by approval, and it can also achieve coverage of more than 100 cities in terms of city coverage.
Tesla is also not using high-precision maps, but it uses visual crowdsourced maps to provide prior information, visual crowdsourcing has previously been a gray area in China, with the advancement of this round of network security reviews, crowdsourcing maps have become more uncertain.
At present, noH has customized an SD Pro map to a certain map dealer, and can also provide transcendental information including lane line top relationship, which we can understand as a set of 2D navigation maps with lane line information.
This set of maps is the basis for vehicles to achieve navigation path planning, but to achieve accurate navigation to choose the optimal route, perform lane change, acceleration and deceleration commands, but also need to perceive and make decision control in the "real skill".
Heavy Sensing is more demanding on hardware
First of all, in terms of hardware, the sensor program is equipped with 3 forward 8 million cameras, 1 positive and rear 8 million cameras, and the left and right side front and side rear are 8 million wide-angle cameras.
In addition, there are two MEMS lidar M1 of Sagitar Juchuang on the left and right sides of the body, and the M1 is our old friend, but it is installed in this position, and it is the first time to see it.
Other home lidars are as far ahead as possible, and even the roof of the car to increase the detection distance, or through the splicing of multiple FOVs to achieve a larger horizontal field of view.
At the end of the day, this set of schemes is more Versailles, there is lidar, hey I just don't sweep forward, is to play (used as a blindfold for the left and right sides). This largely shows the confidence of The End of the Eye in its own visual algorithms, relying on vision to perceive the depth of the forward environment.
At this installation location, I guess the corresponding functional requirement is to make the decision more timely and safer when the system performs a left-to-right lane change. In fact, it is very reasonable to think about it, we treat forward obstacles, the demand for processing is generally not so urgent, basically deceleration and acceleration, and 8 million cameras have more advantages than the current stage of lidar in the detection distance, can detect obstacles at a longer distance, and process more flexibly.
However, in the longitudinal treatment, it is often urgent work, for example, when we avoid the front car, we need to accurately know whether there is a rapidly approaching vehicle in the left rear, and our avoidance and lane change are not safe in the end.
In addition, it is also equipped with the self-developed "Little Magic Box 3.0", which hash rate of 360Tops, and can be upgraded to the four-chip 1440TOPS hashrate version in the later stage, which is enough to meet the computing power requirements of higher-level functions.
"Heavy sensing" requires more algorithm capabilities
The mana is divided into perceptual intelligence and cognitive intelligence, perceptual intelligence is what we often call recognition, prediction, and fusion, and cognitive intelligence roughly corresponds to what we often say about regulation and control plus model training.
First of all, in terms of perception, at the last AI Day, the problem that needs to be solved urgently was proposed: "the perception of time is not continuous, and the perception of space is fragmented."
In just nearly 4 months, a Transformer-based solution was given. Transformer is closer to the way humans understand the world, and can propose a unified pointer to pull together the cross-departmental collaboration of various departments (sensors), the pattern is open, and it comes with a global vision.
On AI Day, Zhixing introduced the self-developed "BEV Transfomer" algorithm, which can take advantage of transformer to stitch together the cameras scattered around the front, back, and left and right of the car and transform them into an overlook perspective, which is equivalent to the God perspective we use when playing racing games.
Friends who have played racing games should be able to feel that from God's perspective, we will be flexible in making decisions and judgments, and the information we have is more direct and complete, although the sense of substitution is almost poor, but it is easier to run good results.
An example of lane line detection is given to illustrate the advantages of BEV Transfomer: more stable and more complete information. In addition, at this stage, our use of lidar perception results is also translated into a God perspective (BEV) through the PointPillar algorithm.
Therefore, if vision and lidar point clouds are mapped in the same axis, and the use of Transfomer's own social fusion capabilities is theoretically more convenient to achieve real lidar, camera front fusion perception, then the system's perception of the environment will have a qualitative leap.
Training is the focus of smart driving
After so many years of development, intelligent driving has more and more algorithms in perception, many of which are still open source, and decide whether they can be used well, how to use well, and also need a lot of data training to optimize the model.