He Xiaopeng said at the 2022 China Electric Vehicle 100 Forum that the city-level NGP effect pushed out within this year will be considerably better than the Tesla FSD effect.
Some people will think that this is another big talk by the CEO of the new power brand, and some people think that this is entirely possible. Today, I come to analyze with you from multiple perspectives whether Xiaopeng's urban NGP effect will be better than Tesla's FSD.
Swing from Tesla across the road to see hardware redundancy
The new domestic power brand represented by Xiaopeng looks much more secure than Tesla in the way of vision + multi-sensor + high-precision map, because it is very critical that the hardware redundancy of Xiaopeng City NGP is stronger than that of Tesla FSD.
It is often said that don't trust your eyes too much, because the world your eyes see is not necessarily real. Indeed, what the camera sees is not necessarily completely true, millimeter-wave radar and lidar can cooperate with the camera for more complete and comprehensive perception, especially when a part of the perception device is interfered with, or obscured, the remaining sensors can continue to complete the perception task, which is the redundancy of perception.
When Tesla's "eyes" are obscured or affected by changes in light and shade, intelligent driving will decisively withdraw and let people manipulate it. Some car owners said: After turning on assisted driving on the highway, after the vehicle enters the tunnel, Tesla suddenly brakes sharply, and then returns to the normal automatic driving state, which is caused by the sharp change in light and the misjudgment of the vehicle. The camera will be like the human eye, in the face of drastic changes in light and shade, there will be a momentary perception failure, which takes time to adapt.
The problem will also appear in rain and snow weather, "God covered your eyes for you, forgot to open", you can imagine what kind of operation Tesla will do. As for the lack of light at night, the problem will become more serious, there are videos showing Tesla in the night ramp fork area swinging from side to side, and there is a situation of entering the wrong intersection, which is the natural drawback of the visual solution in the dim light environment, and it seems that it will not be able to be solved well in the future.
In this regard, due to the high redundancy of the sensor, take the Xiaopeng P5 as an example: it is equipped with two wedge-shaped prism lidar, 12 ultrasonic radars, 5 millimeter-wave radars, 13 high-perception cameras, and a sub-meter-level high-precision positioning unit, a total of 33 sensors. In contrast, the current Tesla Model 3: 8 cameras, 12 ultrasonic radars, and 1 millimeter wave radar. Xiaopeng P5 has 4 millimeter wave radar and two lidars more than Tesla Model 3, one of Tesla's millimeter wave radars is arranged on the front of the car, Xiaopeng's two laser radars are arranged on the front of the car, and the remaining millimeter wave radar is arranged on the front and side of the car, and the radar with the camera can help xiaopeng P5 change lanes more perfectly.
Tesla for the side of the rear of the vehicle quickly up the vehicle, there is a certain perception of the hidden danger, its one of the close-range wide-angle camera coverage 50 m, two mid-range rear side view camera coverage of 100 m, but monocular vision for depth recognition inherently defective, it is understood that the current use of the GIF multi-frame position differential processing, but also can not do good recognition, coupled with the night backward glare of the impact, the rear side recognition perception problem may be exposed more seriously.
The way of relying on the side of the vision and not installing millimeter-wave radar has caused our other domestic new force to increase the range SUV repeatedly in the process of automatic lane change, and the chip shortage was encountered when the model was changed, but it also gritted its teeth to match, which is obviously more complete.
In Chinese cities, what redundancy is more important?
The redundancy of the hardware is strong, and it can also make the perception ability stronger. In cities, stronger perception is absolutely necessary, and the more accurate and comprehensive the perception, the more directly means safer.
Because the algorithm needs to process the less comprehensive information brought by the camera and a millimeter-wave radar, there is no more sensor to assist it, resulting in insufficient perception in the city. At present, the Tesla FSD version is common in the city:
On streets with wider lane lines, snake-like driving occurs and stability is not enough. Social vehicles that are driving illegally will be misjudged, resulting in withdrawal from FSD. On streets with particularly large traffic, vehicles will repeatedly judge, hesitate for too long, and even automatically exit the FSD, which indirectly proves the existence of the upper limit of the footfall threshold. Unprotected turns, sometimes hesitant and sometimes aggressive, resulting in a lack of confidence in driving.
We can see the current strength performance of urban NGP in the Xiaopeng P5 city NGP engineering version test video released by Xiaopeng, especially for unprotected left turn right turn, avoidance of static vehicles in front, these are very difficult to calibrate the situation, P5 can be easily faced, you can see the current urban NGP but strength. At the same time, according to He Xiaopeng, in the test area of Guangzhou City, the Xiaopeng P5 test vehicle equipped with the city NGP has repeatedly maintained that it does not need to take over for tens of kilometers, or only takes over once, which is very close to the effect of high-speed NGP. Previously, the success rate of high-speed NGP through ramps, tunnels and lane changes was more than 90%, which shows that the current urban NGP has achieved initial results. But from the video, it is not all so perfect, and the operation of the P5 also makes people sweat when facing the most Chinese characteristics of electric vehicles suddenly crossing the zebra crossing.
In view of the current situation of these roads with Chinese characteristics, the intelligent driving assistance system of China's local new forces to build cars is based on these situations, and the localization problem is also an important part of the NGP effect to be stronger than FSD.
The head of Baidu's autonomous driving research and development department has said that the complexity of Beijing's road environment is more than 15 times that of California, Beijing is just a microcosm of China's complex road environment, if you go to 3D Chongqing, it is another kind of "complexity", growing up and China's local new forces intelligent driving system, it is indeed more difficult than Tesla born in California, a variety of uncivilized driving behaviors in the city, "different forms" of non-motorized vehicles and "magical" operations, and pedestrians bring you "unexpected", Are the local intelligent driving to play a birth to consider the problem, these problems domestic new forces have spent a lot of time to overcome, do need a large number of sensors to cooperate, self-learning and more in-depth perception, to achieve more redundancy Tesla in the face of these "do not know" objects or situations, and there is no radar sensor, what will it do?
FSD just can't work? Want to ditch the pure visual scheme?
Here, I also want to output some other views, Xiaopeng NGP is to rely on high-precision maps, such as if there is no high-precision map coverage of the place, it will take the initiative to withdraw from the NGP, if it is to some places that are being demolished and built, the map has not had time to update, NGP can not be achieved. In this regard, it can better reflect the universality of Tesla FSD, and the visual solution can adapt to the changing environment and complete intelligent driving through self-learning ability. Of course, at this stage, it is still a bit ahead of its time to consider such a problem, after all, it is still the stage of man-machine co-driving and manned driving. But if it comes to the driverless stage and the steering wheel is canceled, wait for the trailer. Therefore, to complete fully autonomous driving in various situations, visual solutions are still a trend.
The factors that make Tesla so fond of vision solutions are as follows:
First of all, in the process of intelligent driving development, "ghost brakes" and "ghost brakes" continue to appear, millimeter wave radar may produce misjudgments, resulting in contradictions between vision and radar feedback, simply cancel, improve the visual algorithm, once and for all.
Second: Tesla has a powerful set of pure visual ranging methods to replace the responsibilities of millimeter-wave radar. Tesla makes full use of the driving data of a large number of its vehicles for deep learning, making the AI system smarter and more proud to make judgments. I think Tesla's solution is very close to the real driver's control logic: the camera is the driver's eyes, the information seen and perceived is summarized into the algorithm platform, through the "brain" calculation thinking, and then the system according to the experience learned from the data to make response actions. Humans are not like "bats" who can emit ultrasonic waves, so millimeter-wave radar is omitted.
Moreover: the cost problem is a major consideration of Tesla's car manufacturing, Tesla currently uses a single camera average cost of about 150 yuan, the domestic new forces sought after the lidar minimum also needs about 200 US dollars, for Tesla, which is extremely sensitive to upstream and downstream raw material prices, the huge difference in this part of the cost must be reflected in the price.
City NGP sit back and relax? The "wolf" is growing rapidly
Finally, Xiaopeng City NGP can be better than Tesla FSD at this stage, but more and more news indicates that Tesla's new generation of FSD is coming.
The current generation of Tesla FSD (HW3.0), the image sensor is ON Semiconductor's AR0136AT, which is a 2015 product with only 1.23 megapixels. The new generation of FSD (HW4.0) is expected to be installed first on the Cybertruck, which is constantly delaying mass production, and one of the biggest changes is that the image sensor has changed from ON Semiconductor's AR0136AT to Sony's IMX490, increasing the number of pixels to 5.43 million. For every doubling of pixels, the computing power should be increased by 3 times, and the computing power of the FSD chip may break through 1000Tops.
At the same time, the latest production of Model S and X in February this year really canceled the millimeter wave radar, Tesla is really going to "stumble to the end" on the road of vision, and Xiaopeng City NGP is about to usher in a new round of challenges from FSD.