Tesla is so big, can’t it accommodate a small radar?
Latest update time：2022-10-15
Reads：
Although Tesla AI Day has only been held twice, it is also called the Spring Festival Gala in the field of autonomous driving in the industry because it contains so much practical information and the technology is so hardcore.
Last year, the key words of AI Day were BEV, Transfomer and Hydranet. This year, it has become Occupancy Network. The technology iteration capability is amazing. No matter how the noun changes, Tesla’s insistence on the pure visual route has not wavered at all.
Reflected in reality, as the AI neural network architecture becomes more and more mature and the cloud supercomputing DOJO is launched, the hardware required by Tesla in the perception process can be "simplified from complexity" and become 100% pure vision. Therefore, Musk is also known as the "Radar Killer".
In 2018, Musk mocked LiDAR as a "crutch", "useless ribs" and "appendix" for autonomous driving; in 2019, Musk made another judgment, "Those who rely on LiDAR are doomed to fail." In 2020, when Xiaopeng announced that it would launch the P5 with LiDAR, Musk even vented his anger on Twitter. Later, He Xiaopeng responded strongly on Weibo, claiming that he would beat Tesla to a pulp.
In 2021, Tesla became more radical. In May of that year, it announced that it would remove the millimeter-wave radar from the newly produced Model 3/Y for the North American market, which caused a lot of controversy in the industry. Last week, Tesla turned its attention to the ultrasonic radar.
The question is, when the entire industry is trying to install more and more powerful radars into cars to provide multi-sensor guarantee for autonomous driving, why does Tesla insist on not even using a single radar? Is it just because of the "money" issue?
In this article we try to answer several questions
1. Why does Tesla keep abandoning radar?
2. What is Tesla’s confidence in removing various radars?
3. Is there still a chance for radar to be installed on Tesla?
0
1
The work is too rough and the price is too high
The reason why Tesla abandoned radar is not complicated. To sum up, it is either that the work is too rough or the price is too high. The two key words of Tesla's official certification are:
signal-to-noise ratio (ratio of signal and noise) and cost
.
The millimeter-wave radar that was first cancelled by Tesla was cancelled because
the signal-to-noise ratio was too low
.
The working principle of millimeter wave radar is very simple. It emits radio waves, receives echoes, and then measures distance and speed based on time difference and phase difference.
But the problem is that mainstream automotive millimeter-wave radars have low resolution, lack height measurement capabilities, and are accompanied by clutter interference throughout the process, which causes objects to often appear as a single point in the detection results (or there may be no point at all, or there may be clutter that should not be there), making it difficult to determine the shape and category.
This is like asking a person to cover his eyes and use only one finger to feel and judge
the properties of the object
in front of him
. The effect can be imagined.
Most of the use of millimeter-wave radar in the automotive industry is to maximize its strengths and avoid weaknesses, focusing only on its ability to track dynamic objects.
But static objects are a nightmare it cannot escape. In order to prevent millimeter-wave radar from treating various stationary objects such as manhole covers and overpasses as obstacles, the industry will basically block its detection signals for stationary objects, otherwise it will lead to "ghost braking" The situation happens frequently.
Therefore, in traditional automotive millimeter-wave radars,
a complex real world is often reduced to a plane with only some points.
The world through the eyes of millimeter-wave radar
Moreover, the millimeter-wave radar used by Tesla has always been the entry-level product ARS 410 launched by China in 2016, and its performance is outdated. In Tesla's technical framework in the past two years, the detection signal of millimeter wave radar often becomes "noise" in the perception system, contaminating the perception data, misleading other sensors, and may increase unexpected situations.
With this preparation, it seems natural that Tesla discontinued ultrasonic radar this year.
To some extent, ultrasonic radar is a degraded version of millimeter-wave radar. It does not have the functions of millimeter-wave radar, and it has all the problems of millimeter-wave radar. Its detection distance is short (usually within 3 meters), it cannot measure speed, and it cannot measure the outline of objects. Its main advantage is that it is cheap (tens of yuan per piece, and a complete set of solutions is only a few hundred yuan), and it can be used in low-speed scenarios to assist in collision avoidance when reversing and parking.
Compared with millimeter-wave and ultrasonic radars with weak performance, lidar has greatly improved the signal-to-noise ratio problem, so much so that Tesla has quietly used this "crutch".
Tesla, which sneered at LiDAR on the surface, actually installed LiDAR on a batch of vehicles in 2021. The purpose was to use the latter's highly accurate measurements to help calibrate the visual algorithm. However, these vehicles were only test models, and at best, LiDAR only played a temporary role in Tesla for a period of time.
Model Y equipped with lidar test
LiDAR has never been used in Tesla's mass-produced cars, mainly because of another fatal flaw:
the cost is too high
.
The concept of cost killer is engraved in Tesla's DNA. When Tesla started its business, it chose 18650 cylindrical batteries as its power battery because of its mature technology and low price. A newer example is that Tesla took the lead in using the rear axle integrated die-casting process on the Model Y white body, integrating the original 70 parts into one piece, which not only reduced the weight but also cut the cost by 40%[1].
As the number one manufacturer of smart electric vehicles, the high cost of lidar makes it difficult for Tesla to accept it.
At present, the price of a high-performance lidar can reach tens of thousands of yuan, and the cheap one costs 3,000 yuan [2], while the price of a high-definition car camera module is only 400 yuan [3]. Tesla’s self-developed FSD chip The estimated cost is only $200.
If it is equipped with lidar, the cost of Tesla's smart driving hardware may directly double, which runs counter to Tesla's vision of "technological inclusion".
In addition to the two major issues of signal-to-noise ratio and cost, radars have other problems: different data formats, high technical barriers for calibration and fusion, and high computing costs; different operating frequencies (10 frames per second for lidar and 36 frames per second for cameras), and data time synchronization are also technical barriers. Automakers that adopt the multi-sensor route often have to support teams of hundreds or even thousands of people.
Tesla's desperate move of cutting out all radars will allow it to fight a last-ditch battle and focus all its funds and resources on the technology route that Musk trusts the most - pure visual autonomous driving.
0
2
Vision "Kills" Radar
In the past few years, the reason why Tesla rejected radar was, after all, that they were not reducing costs and increasing efficiency fast enough. The B-side of the story is that the purely visual technology that Tesla has bet on has advanced by leaps and bounds, so much so that in less than three years, it has replaced various radars.
In fact, Tesla's early Autopilot took a multi-sensor route. In theory, the camera and millimeter-wave radar worked together to achieve assisted driving.
However, when faced with complex situations, the two often give up each other - the millimeter-wave radar is like a blind man groping in the dark in the horizontal direction, while the camera is struggling to identify vertical images frame by frame, just like two two-dimensional creatures arguing about whether the world should be horizontal or vertical.
Upper picture: millimeter wave radar field of view; lower picture: camera field of view
This period was also the peak period for Tesla's assisted driving accidents. In 2016, a Tesla crashed into a white truck that had overturned and has long been a classic example of the failure of multi-sensor collaboration: the millimeter-wave radar failed to detect the elephant, the camera turned a blind eye to the truck with blurred features, and the human driver did not provide a backup plan, so the accident was inevitable.
The change began in 2020, when Tesla realized that
it was difficult to accurately restore the three-dimensional world by relying on two two-dimensional creatures to exchange information
.
That year, Tesla launched FSD beta, fully turning to cameras with higher potential to reconstruct the intelligent driving algorithm: on the one hand, through deep learning training, the visual algorithm learns to estimate the distance and speed of objects, and obtains preliminary The three-dimensional perception ability; on the other hand, the information obtained by multiple cameras is fused from a bird's-eye view, and the time dimension is added, so that the algorithm "lives" in a four-dimensional space-time that is closer to reality.
Speed and distance measurement are the foundation of millimeter-wave radar, which means that
the day when Tesla
's
visual ability
is launched is the day when millimeter-wave radar will be out of work.
In May 2021, the millimeter-wave radars on Tesla’s North American models were all destroyed. In June of that year, Andrej Karpathy, Tesla’s director of artificial intelligence (who left the company this year), explained that in actual tests, the speed and distance measurement capabilities of visual algorithms have approached or even surpassed those of millimeter-wave radars[4].
This year, Tesla's pure vision algorithm has gone a step further. The emergence of the Ocuppancy Network has allowed the camera to target a more powerful opponent - LiDAR.
The purely visual occupancy network algorithm divides the perception space into three-dimensional grids. By detecting whether the grids are occupied, it can measure the volume of objects in a low-computing-power-cost, low-computational-latency manner, including all kinds of special-shaped objects that have given headaches to smart driving teams around the world.
This solves the classic problem of visual perception: "if an object is not recognized, it does not exist."
Before this, the industry generally believed that expensive LiDAR was the correct answer.
Tesla used the network to accurately detect the extremely long bus that "moves forward and remains silent". This kind of target is usually a problem in the industry.
After the vision algorithm has the ability to benchmark lidar, the weak ultrasonic radar has logically lost its necessity to exist. Cutting it off will save Tesla hundreds of dollars per vehicle.
Tesla officials stated that after the ultrasonic radar goes away, the occupied network will take over its work: with the algorithm update, Tesla's difficult-to-use automatic parking capabilities will be enhanced, and the smart summons that have been bouncing will also follow.
So far, in the horse race between cameras and radars in the Tesla system, the camera has completed the simulation of radar capabilities through the evolution of algorithms, achieving a "
visual radar"
effect at a low cost, one of the three. Faced with this epic struggle, the radars can only graduate one after another.
However, some radars may become history, and some radars are expected to return to work at Tesla after transformation.
0
3
Millimeter wave radar re-employment
The first to be laid off and re-employed is most likely to be millimeter-wave radar.
In June and September this year, Tesla was found to have passed the certification application for two self-developed millimeter wave radars submitted to the FCC (U.S. Federal Communications Commission). Green, an American white passenger, also found it in the new bill of materials of Tesla models. The figure of millimeter wave radar.
Tesla applies for certification of self-developed millimeter wave radar
The question is, Tesla just eliminated the millimeter-wave radar last year, why would it slap itself in the face with the speed of light again?
One answer may be
privacy and data security
.
In 2019, Tesla launched Sentry Mode and Pet Mode. These two functions use cameras to continuously monitor the environment inside and outside the car when the owner leaves, to prevent theft and protect pets in the car. However, after the function was launched, it has aroused the vigilance of the Norwegian military, Berlin police, and the Chinese government on the grounds that high-definition cameras may cause privacy leaks or national security risks.
The low-resolution characteristics of traditional millimeter-wave radar are actually an advantage - it can realize these functions without excessively collecting high-precision data.
For example, the 60 GHz in-car radar that Tesla applied for has several potential uses that overlap with the Sentry/Pet Mode, according to the materials submitted by Tesla: monitoring vital signs in the car, which can be used for child sensing (to prevent children from being forgotten in the car); sensing within two meters outside the car/broken window detection for theft prevention; and performing gesture recognition to enrich the forms of in-car interaction [5].
Therefore, this type of millimeter-wave radar may help Tesla models meet the privacy and data compliance requirements of various countries while launching new features.
However, in the eyes of technology enthusiasts, the highlight is another millimeter-wave radar that Tesla has applied for confidentiality.
Because the confidentiality order was not lifted until December this year, its specific parameters and purpose are still unclear, but from the test reports that have been made public, it is known that this is a 77Ghz radar with a 6-receive and 8-transmit antenna configuration. This information points to its identity as a more advanced millimeter-wave radar - 4D millimeter-wave radar, also called imaging radar[6].
Tesla removed traditional millimeter-wave radar from its cars largely because its resolution was too low to accurately detect and identify objects. But this problem is not unsolvable. Generally speaking, the higher the operating frequency and the more transceiver channels, the higher the resolution of the millimeter-wave radar.
The high-precision perception requirements of autonomous driving are forcing the continuous evolution of the accuracy of millimeter-wave radar. High-resolution radar has long been a popular technology, and this year is called the first year of 4D millimeter-wave radar being put on the market.
In China, SAIC Feifan R7 is equipped with 4D millimeter-wave radar from ZF, which has significantly improved the intelligent driving effect - compared with traditional millimeter-wave radar, which can only generate a small number of points on a plane, 4D millimeter-wave radar can draw three-dimensional space. The point cloud image in , which is similar to the lidar effect.
Arbe 4D millimeter wave radar prototype effect, the company has had technical cooperation with Tesla
In fact, 4D millimeter wave radar is a sensor between traditional millimeter wave radar and lidar. Its perception accuracy is significantly better than the former, and the cost can be as low as 1/10 of the latter. Tesla has always liked such a cost-effective choice. It is logically reasonable for millimeter wave radar to return to Tesla.
At present, although Tesla's pure visual intelligent driving algorithm has made great progress, its speed and distance measurement are still "estimates" based on deep learning, and its reliability will be greatly reduced in the face of weather conditions such as rain, snow, and fog.
The speed and distance measurement of millimeter-wave radar is based on the calculation of physical principles and has all-weather detection capabilities. A 4D millimeter-wave radar with a significantly improved signal-to-noise ratio can complement the capabilities of Tesla's cameras and improve system performance.
In fact, Musk has never been completely uncritical of millimeter-wave radar. Earlier this year, when interacting on Twitter, he left a rather suggestive reply: "
High-precision (millimeter-wave) radar is the right answer.
"
0
4
end
Tesla’s repeated jumps in its attitude towards radar may make some melon-eaters a little nervous - didn’t they say that All in vision can be changed at any time?
For a long time, the public has had some misunderstandings about Musk's "first principles". They believe that the performance of "first principles" in Tesla is to imitate humans. People can drive by relying only on vision, so autonomous driving should also be the same.
In fact, Tesla chose a vision-based intelligent driving solution. The first reason is that it can see richer information (such as color, semantics, which Musk calls high-order "qubits"), and the second reason is that it is low-cost. . The development of technology is dynamic. As long as the above conditions are achieved or approached, different sensors will find their own uses.
Tesla's ambiguous attitude towards radar can also reflect the company's true first principles - to set a difficult goal and then find the most cost-effective engineering means to achieve it.
(PS: Domestic Tesla model millimeter wave radars and ultrasonic radars are still on the job and have never been unemployed)
References:
[1] Electric vehicle companies are rushing to enter the market. Why is integrated die casting popularized by Tesla? See Zhi Research Pro
[2] I hit an electric car, emptied my wallet, and enjoyed it
[3] Autonomous driving enters the fast lane, and vehicle-mounted cameras are experiencing both volume and price increases, Orient Securities
[4] CVPR 2021 Workshop on Autonomous Driving, Andrej Karpathy
[5] Federal Communications Commission DA 21-407
[6] Tesla embraces radar again? Fanzhi Grocery Store
-END-
The content of this article is reproduced from: Yuanchuan Automobile Review, for communication and learning purposes only. If you have any questions, please contact us at info@gsi24.com.
▼ Highlights from previous issues ▼
The parallel-imported iPhone 4th-hand business is declining in Huaqiangbei
Made in China stuffed iPhone 14
just! Russian air strike affects Samsung building in Ukraine
What happened to semiconductors?
A must-read for chip people | Understanding the chip industry in five minutes from three dimensions
京公网安备 11010802033920号