Author: Mark
Producer: Redplanx
Head image: Daydreamer promotional image
A few snowflakes drifted slightly from the sky in Beijing.
Today, it is a great honor to have four outstanding representatives in the field of visual perception, namely Maxieye, Minieye, Magic Vision and Polar Intelligence, who welcome them to the "Dreamer" program.
Although there are many of us, try to discuss some simple questions, not too esoteric, the host said.
1. Cross-border
I learned that you were all in the field of commercial vehicles before, doing assisted driving, also known as commercial vehicle double warning, why have you crossed over to the passenger car market?
Minieye: The war in the commercial vehicle market, 2020 has basically been fought, the market share is more obvious, who is the customer is also relatively clear, the next thing to do is to serve the existing customer well, from attack to defense.
Maxieye: I don't think so, but I'm trying to say that the first phase of the commercial vehicle war may have just been completed, but it's not completely over. Including us, and a few of you here are still fighting in this market.
In 2020, our shipments in the commercial vehicle market will be more than 100,000 units, and the shipments in 2021 will basically double, and will continue to maintain rapid growth this year.
Extreme Intelligence: I agree with Maxieye's statement, in fact, the commercial vehicle war is not completely over, but everyone pursues a larger market, so it will cross over to the passenger car market.
More of our shipments are in the aftermarket, of course, there is also a front-loading market, but the volume is relatively small, last year's shipments in the scale of more than 100,000 sets.
The front-loading market is relatively easy to ship on a large scale, a customer may have tens of thousands of sets of scale, after-loading market we need to talk with insurance companies, dangerous chemical transportation companies, logistics fleet companies and bus fleet companies, one by one.
At present, our company's vision-based DMS products are also exclusively supplied to the Xiaoma Zhixing L4 vehicle, Tucson Future and Didi autonomous vehicles, which can also show that everyone recognizes us.
Moshi: We are similar to Jimu Intelligence, and the shipment volume in the aftermarket is relatively large. It mainly includes stock vehicles such as buses, muck trucks and hazardous chemical vehicles, providing them with forward-looking ADAS, DSM (fatigue driving) and BSD (blind spot monitoring) and other products for applications.
(Magic Vision APA test image)
However, in the passenger car assisted driving market, we have signed cooperation with the top 9 OEMs in China, the main layout is APA and AVP, we have been doing parking products since 2018, and it has been more than 3 years now.
I have a question, the host said, since you have been in the commercial vehicle market for several years, why don't you just go up the commercial vehicle and go from L2 to L4? Because of the single driving route and scene (mainly on the highway), commercial vehicles are not considered the most valuable and the best to land?
Magic Vision: Commercial vehicles are actually not as easy to do as everyone thinks. The main engine factory is very strict about the cost control, and their maximum willingness to pay for a security function product is about 1,000 yuan. So you let everyone spend 1,000 yuan to install a double warning, there is no problem, but you want everyone to spend 5,000 yuan to install an AEB, in fact, many commercial vehicle OEMs are not very happy, so the domestic commercial vehicle AEB market has not started.
Secondly, the two big brothers in the field of commercial vehicle wire control, WABCO and Knorr-Bremse, have always monopolized the wire control technology of commercial vehicles, although we have completed a large number of docking with the international Tier1, from the perspective of the wire control chassis, it will not become a bottleneck restricting mass production.
But always send people under the fence, and if one day these 2 big brothers also release auxiliary driving products, the adaptation of wire control is definitely better than ours, and we basically can't play.
Jimu Intelligence: Yes, you see very few companies in China doing advanced assisted driving of commercial vehicles, why? Although the driving environment of commercial vehicles is relatively single, it is far behind passenger cars from all angles such as perception, planning decision-making, and wire control.
Do longitudinal control can also be, once involved in horizontal control, your perceived distance, vehicle positioning and trajectory prediction, in fact, are difficult to passenger cars, to put it bluntly is more difficult to upgrade the dimension, so everyone chooses to change a battlefield.
In the field of passenger cars, as well as benchmarking companies such as Tesla and Wei Xiaoli, at least in terms of education market and user acceptance, everyone has a little feeling; in the field of commercial vehicles, there is no benchmarking enterprise, it is too difficult to promote, and user education lags far behind the passenger car market.
Regulations have also been slow to open up, unlike in the United States, where more than forty states allow unmanned commercial vehicle testing and operation. There are very few test roads in China today.
As we all know, this wave of auxiliary driving boom in commercial vehicles is actually driven by regulations, requiring the operation of buses in 2018 to install double warnings, 2019 to require the operation of buses to install AEB, and 2021 to require trucks to install AEB.
One of the characteristics of the regulation is that it is fast, but it will also produce some sequelae, and sometimes it will limit the development of the market and make end users resent this thing. After the promotion of regulations, it may leave a mess, and it may be dismantled and re-engaged in later stages.
From several aspects, it is too difficult for commercial vehicles to upgrade dimensions, or it is impossible to go from L2 to L4, so everyone has turned to passenger cars to upgrade the dimension road.
Maxieye: As Jimu said before, the chassis technology and middleware of commercial vehicles are firmly in the hands of international parts giants, and the ability of domestic ADAS manufacturers to adapt to the chassis needs to be improved.
Previously, when domestic OEMs faced Mobileye, they either accepted it or had no products available, because large suppliers would not care about some of the buyer's needs. This is very detrimental to building an autonomous and controllable supply chain, so we have invested in the field of visual perception to make up for China's shortcomings in this regard.
2. ADAS Marketplace
Oh, that's the way it is, the host said, what do you think of the passenger car ADAS market?
Minieye: The next three years must be the golden period of passenger car ADAS development, and from the fact that several of us here are vigorously deploying, we can see how much we attach importance to the passenger car ADAS market.
(Minieye Visual Perception Products)
The commercial vehicle ADAS business is our first gold in autonomous driving. Judging from the number of passenger car projects we got this year and the timing of mass production, the first year of autonomous driving has really arrived. Over the next 1-3 years, the passenger car ADAS market will accelerate into its formative years.
Extreme Intelligence: I'm not so optimistic, and from what I know, I think it's good that advanced assisted driving can all get on the bus in the next 5-10 years.
From the results, the landing application of ADAS is far less rapid than the industry previously expected, and everyone underestimates the difficulty of the technology in the actual landing process. A lot of people talk about ADAS, but how many people use ADAS, and how much the onboarding rate of ADAS is, I believe it's very low.
The main reason is the user's expectations of ADAS, and the contradiction between the user's actual experience, when experiencing, sometimes it is really not valuable to the user, more is interfering or in the chaos, including some brand manufacturers to make products can not be separated from such a situation.
We cooperated with Hyundai in 2011, but considering that the pre-installation cycle is very long, it takes about 5 years to support, and at that time, the OEMs are not very important, so we finally decided to do it ourselves. Now, if you cooperate with the main engine factory, it will take at least 3 years to mass production.
Magic Vision (laughing): I have a middle attitude. We used to talk to the OEMs with our heads down, and our status was very low. Now it's better, some OEMs take the initiative to find us to cooperate. We began to engage in parking in 2018, Minieye began to engage in smart cockpits in 2017, and has not yet been rolled out on a large scale, and we have psychological expectations for this time.
This great separation between expectation and reality has all been experienced. In terms of automatic and autonomous parking, it is a gradual development process in itself. The cost of assisted driving in the main engine factory is also constantly raising psychological expectations, before 1,000 yuan is a top, and now about 10,000 yuan can also be accepted.
Maxieye: Indeed, in terms of cost, a set of autonomous driving systems may cost hundreds of thousands, and assisted driving systems may cost up to ten thousand yuan, and the gap between them needs to be slowly adjusted.
Unlike the commercial vehicle market, passenger cars are not regulated, but are fully marketed. This determines that this market may be slower, not a business model that can be launched on a large scale in 2-3 years and can be well verified.
Another point is that the reason why we switched from commercial vehicles to passenger cars is because before we may have mainly done double warning, without involving planning and control modules (plus AEB is too difficult to implement), now we are moving to passenger cars, just to improve the ability of planning control modules.
3. L4 VS L2
The host laughed and said that in fact, this kind of debate is very good, if everyone rushes to look at a direction, it will be a problem. How do you see L4's dimensionality reduction?
Maxieye: Actually, we're not afraid to compete with L4. L4's company is through continuous financing, in the next five to ten years, with investors' money to engage in research and development; we are making money while landing, while growing.
I think there are several fundamental problems that make it impossible for us and L4 companies to compete. The object of our service is the main engine factory, and the object of the L4 company is the travel company, we are bound with the main engine factory, and the L4 company and the travel company are bound.
So our main competitor is a foreign Tier 1 company, not an L4 company. L4's competitors include mobility companies, which can also be understood as a competition between us and Tier 1 and OEMs. Like Bosch, he employs 400,000 people worldwide, which is larger than any oem in China, and he can do every part of the car, including the car production line, but he just doesn't do the final integration, but he does the Tier 1 business.
Our competitors are Bosch, Continental, while L4's competitors include not only mobility companies, but also OEMs. Because L4 companies also end up doing integration, you can understand that the core of the autonomous driving company is not the algorithm, but the multi-threaded, multi-process communication and computing resource scheduling, which is an engineering problem.
OEMs and L4 companies may end up becoming system integrators.
So the competitive relationship scores clearly. L4's ultimate way out must be travel services, and our ultimate service must be to help car companies sell cars better. So I think that there is no competition at all, whether from their own positioning or from their ultimate goals.
On the other hand, the cost of L4 and the cost of L2 are fundamentally different. For L4 companies, they must add lidar, but a semi-solid-state lidar is the cheapest and costs $2,000, which OEMs and consumers can't afford at this stage.
For OEMs, a security feature they can accept is a thousand dollars. Making a line may not be so important for mass production and cost requirements, but for OEMs, if the price is too expensive, it can't be sold at all, so they think more about mass production, safety and cost.
In the end, L4 companies can only recover the cost of hardware through the way of operation; if it is for the C-end, it is more of an experience, or to put it bluntly, it is to sell a car, and the cost is still a very critical car purchase factor.
(Maxieye Advanced Driver Assist Architecture Diagram)
In particular, some L4 companies have also begun to engage in assisted driving, I think it is positioning and strategy is not clear, many people think that from L4 to L2 is a horse Pingchuan, but in fact, the threshold in this field is very high, not high in technology, more is to test the engineering ability, channel response ability, fragmentation needs to solve the ability.
Moshi: Indeed, we know how to meet the needs of users while keeping costs low, and help him make more detailed improvements. But L4 has a hard time doing that. To take a simple example, 5R1V is a scenario where they simply don't know how to do it because the data schema is completely different.
L4 down to advanced assisted driving usually uses a more luxurious sensor lineup, considering the main engine factory's acceptance of cost, such a configuration can often only be mounted on the high-end model, the amount is actually very small.
Another aspect is the huge labor cost and computing power cost. As a fixed-point cooperative enterprise, we must be stationed on site and debug with the OEMs, because many localized subtle scenes require a lot of manpower investment.
Algorithms and architectures must also be constantly cut and adjusted, and run safely on the computing power of more than a dozen TOPS, which actually tests our patience and ability.
Of course, if they think they can do that, they are welcome to join us and compete with an international Tier 1 like Bosch.
The moderator then asked, "How do you think you can become the new Tier 1 supplier compared to Bosch?"
Minieye: On the one hand, Bosch masters the core technology of the underlying layer of the car, and on the other hand, they also make their own ADAS products, but we are not afraid of them, just like Mobileye.
Bosch is good at making batch products, but once it encounters trivial things that need to be handled, it will seem stretched, which is exactly what we are good at.
Moshi: Yes, the driver assistance system products of foreign parts suppliers are often "unsatisfactory" in the domestic market. There is a consensus in the industry that ADAS systems need to be developed in depth localized to be closer to China's road conditions in terms of perception-decision-execution. This has become an opportunity for many domestic auto parts to overtake in curves.
In addition, at present, the main energy of the international Tier 1 is focused on the joint venture brand, and the investment in domestic independent brand cars is not large.
Extreme Intelligence: I also agree. It is estimated that in 1-2 years, we and their position in the OEM will be on an equal footing. Although there are many software development teams like Bosch, although the amount of investment in software is very large, it is actually slightly insufficient in the ability of a single point, which will be constrained by the entire market and mechanism.
We will cooperate with OEMs in depth in technology and software to overcome some engineering challenges, which Bosch does not have in the short term.
The host nodded, just now the question is a bit serious, let's find some light topics, I know that before starting a business, are experts in the field of vision, can you talk about your industry experience?
Minieye: Our predecessor was an ADAS project done by 6 small partners in the Singapore government, and later discovered this entrepreneurial opportunity. In particular, so far (since 2013), our 6 friends are still together.
Recently I saw a business summary in the circle of friends that is very classic. Vision experts are following Mobileye's steps; those doing optoelectronic research are following Velodyne's steps to do lidar.
Visual perception and lidar perception actually have a lot of similarities. Of course, like Hesai and Sagitar, both financing and valuation are much higher than ours, and it can only be said that they have chosen a good market.
We are also the two most controversial gangs, the perception scheme represented by Tesla and the lidar-based perception scheme represented by Waymo.
In terms of composition, we are the same, we have chosen to embed deep learning algorithms on FPGA chips to do chip-based processing. We may need to moderately reduce the underlying operating system, algorithms and models because of cost and computing power, so that it is lightweight enough to run safely on a limited vehicle-level computing platform.
An AI chip is hundreds of dollars, a semi-solid-state lidar is thousands of dollars, no main engine manufacturer dares to easily get on the car, the cost is too expensive.
Magic Vision: Yes. FPGA for our initial period is indeed very helpful, we have set up a special FPGA team for this purpose, the same is true of the companies here, there is a special team responsible for FPGA software and hardware collaborative research and development, the algorithm engineering landed.
It's not a joke that we're countering Mobileye's development in China with "deep learning + FPGAs." In fact, it is not easy to do the algorithm in the FPGA, many companies have not done it, but we have done it.
With the development of Moore's Law, it is possible to copy the computing power of the X86 architecture on the computer side into the embedded chip.
Maxieye: Yes, yes. Although the FPGA chip is not as stable and secure as the ASIC chip, it must be compared, and it is the general trend to arrange FPGA chips on the domain control computing unit in the future. Although we spend a lot of effort, it also accelerates our calculation speed.
Many people wonder why we don't use binocular cameras, in fact, the cost does not allow. Binocular camera plus computing power, the cost may be more than 2 times the single purpose, the main engine factory will not accept such a cost at all; and the binocular camera is too large, too occupy the front view space, the user will feel depressed, and it is not easy to install.
We adhere to the era of equal rights, so that every user can feel the charm of intelligent driving, after all, high-end models are still a small part.
It's not too early, so I'll ask one last question, what do you think L2 and L4 will end up with, or who will have the advantage in the future?
Minieye: My feeling is that L2+ assisted driving will definitely land on a large scale first, and at present, the OEMs attach great importance to this piece, and every company is doing fixed points.
L4, because of cost and regulations, may not land in the next few years.
Extreme Intelligence: I personally feel uncertain. At present, although there are many OEMs doing fixed points, they are all taking some new models as fixed points, and the real number of models has not actually begun to do fixed points.
In fact, L4 companies are now also enjoying the dividends of L2+ high configuration, and now many models, whether it is cameras, millimeter wave radar, or lidar, have been on the car, which is actually reducing the hardware cost of L4.
(Jimu National Production L2 Intelligent Driving Scheme)
Including our L2+ driver assistance solution, we are trying to integrate the camera with other sensors, such as lidar, to study how to minimize the internal friction of multiple sensor combinations and create a perception system that is more suitable for autonomous driving.
And according to the disclosure of L4 company, about 2023, or 2024 will launch a new generation of perception hardware solutions, at present, the overall cost can basically reach 40-500,000 or so, which includes the price of the car.
So I'm thinking, the current advanced assisted driving basically gets 2023, 2024 will appear in the real sense of L2+ production car, then who is the mainstream, or unknown.
Maxieye: I'm still confident in L2+. During the period from 2016 to 2018, it can be said that it was a sudden spring breeze, and all of a sudden, funds poured into the automatic driving track. But by 2019, it cooled down rapidly.
In the second half of 2020, high-level assisted driving began to be hot again, I think thirty years of Hedong, thirty years of Hexi. Assisted driving will be hot from 2020 to 2023, and the collision of L2 and L4 will begin in 2024.
Who will prevail at that time will have to be seen.
Earlier we said that if we can't keep up with Tesla's pace, then there may be greater challenges in the future. Tesla is now fully visual, should we also keep up? I think Musk is thinking about, on the one hand, his visual perception is really advanced, on the other hand, it is still a cost problem, so he has removed other sensors.
The cost is too expensive, and consumers don't really buy it in the end.
In fact, each route has the advantages of each route, because the technology landing is relatively fast, we can form cash flow faster, if the next 2 to 3 years can achieve IPO, we can leverage more resources.
Magic Vision: I don't think it's a good judgment. We started doing APA in 2018, and by 2022, this function cannot be mass-produced on the car on a large scale; Minieye has started to do dms for passenger cars in 2017, and does smart cockpits, and now it is actually a very low load.
Why, I think DMS needs to ensure two points: the first is available, the second is easy to use, the principle is to miss a little and not to false positives, sometimes users will be very disgusted with the continuous false positives, and finally directly turn off.
So, I think the passenger car ADAS still has a long way to go, and it can't be improved quickly in 1-2 years. I believe that everyone has their own company plan, but they just don't say it.
Some people said that L3 is cannon fodder before, because compared with L4, he is still a man-machine co-driving mode, out of responsibility, the driver still has to carry, not as clear as L2+ responsibility points. But I have a feeling in my heart that L2+ may also become cannon fodder, and we spent a lot of manpower to do this, and it may not be worth the loss in the end.
This is really not easy to judge, I still talk less, so as not to offend the entire auxiliary driver.
Of course, our ADAS vendors are also striving for business upgrading, but they will not be as aggressive as L4 companies, and will take it slowly, first in construction machinery, unmanned port and other scenarios.
The host smiled and said, I can understand the logic you are talking about, but it does make some sense. I also think that advanced assisted driving may not be as fast as everyone expected, and the L4 is not as slow as expected.
Everyone nodded and smiled unspeakable.
Thank you all for taking the time to participate in this show. Okay, that's the end of today's show.
The four good buddies left the studio as they walked and discussed.
-END-