Car stuff (public number: chedongxi)
The author | Wooden rice
Editor| Xiaohan
How far do you think driverless is? 5 years, or 10 years?
This company is trying to get you to buy a driverless car in 2024, the L4 class.
Last year, Yuanrong launched a new autonomous driving solution, which is the lowest cost autonomous driving mass production scheme in the industry, and has the possibility of mass production.
After six months of intense research and development, just yesterday, the team carrying yuanrong's new plan was officially on the road, which means that the era of really letting go and letting the car run on its own is really accelerating.
According to the news revealed by Yuanrong Qixing, this latest L4 level automatic driving solution has landed on the Feifan MARVEL R, Lincoln MKZ, Geely Geometry A and other models, and has begun to cooperate with major car companies to truly put this L4 product into mass production vehicles.
▲The Yuanrong departure L4-level mass production plan has been adapted to models such as the Feifan MARVEL R
Doing L4 level autonomous driving mass production products is indeed relatively rare at this stage.
Che Dong recently had a dialogue with Zhou Guang, CEO of Yuanrong Qixing, who told Che Dong that in fact, from the beginning of its business, Yuanrong Qixing had planned to make L4 level autonomous driving mass production products.
From the end-game goal, to achieve true autonomous driving, then the data collection of L4 autonomous driving is crucial - because no one in the market has done such a thing now, most of the mass production autonomous driving products are still in L2 level, and a small number of them can reach L3 level. But the data collected by all these products is actually not very helpful for achieving mass production L4 autonomous driving.
Therefore, in this sense, the set of L4 level autonomous driving products launched by YuanRong Is very critical - through mass production, let the L4 level automatic driving products get on the car, collect relevant data and then promote, this development path is actually very clear.
First, 30 cars landed in Shenzhen L4 program has the lowest cost in the industry
On the streets of Shenzhen, if you see a self-driving vehicle driving smoothly forward, you may not be too surprised. In the past few years, there have been many unmanned vehicles with "big hats", various cameras, and fully armed vehicles shuttling here.
But what you don't know is that there is a self-driving company that has secretly "disguised" its own self-driving car as a ride-hailing car.
These unmanned vehicles, which look like ride-hailing cars, are actually robotaxis that Yuan Rong started. Just yesterday, Yuanrong Kaixing officially announced that its Robotaxi fleet has welcomed 30 new cars - and this batch of unmanned vehicles is equipped with Yuanrong Qixing's latest L4 level autonomous driving solution for front-loading production, DeepRoute-Driver 2.0.
In the interview, Zhou Guang proudly told Che Dong: "This is the lowest cost mass production level L4 autonomous driving solution in the industry at present. ”
Zhou Guang can say this, in fact, he is very confident. At present, the mainstream L4 automatic driving system in the industry has hundreds of thousands of cars. Among them, only one imported mechanical lidar has even sold for hundreds of thousands, even if it is domestic, it must be tens of thousands, and a set of high-precision positioning combination navigation also needs tens of thousands.
For reference, Baidu previously announced that the L4 level of autonomous vehicle + kit is 480,000. So how much did Yuan Rong start?
Zhou Guang told us that the hardware cost of the entire system is about 10,000 US dollars (about 64,000 yuan), of which about 70% or 80% of the cost is spent on chips and radar.
According to him, the front-loading scheme of the Yuanrong departure is to embed 2 to 5 solid-state lidar and 8 cameras in the body of each car. In the future, its sensor configuration can also be adapted to different sensor configurations according to the design needs of the depot.
Che Dong noted that the L4-level automatic driving scheme started by Yuanrong is equipped with NVIDIA DRIVE Orin vehicle specification-level chips.
▲ Yuanrong started the L4-level mass production scheme using NVIDIA Orin chip
This is very interesting, because most car companies and autonomous driving companies only take NVIDIA's Orin chip to do L2-level automatic driving solutions, after all, the computing power is limited. So how did YuanRong Qixing achieve the L4-level mass production automatic driving solution by relying on the chip with small computing power?
Zhou Guang said that this is based on yuanrong qixing's self-developed reasoning engine technology to achieve such an effect.
When talking about this year's planning, Zhou Guang told Che dong that the total number of Robotaxi teams starting from Yuanrong is about 150 units, most of which are operating in the Shenzhen base camp, and have achieved arbitrary point-to-point operations in the shenzhen operating area.
The DeepRoute-Driver 2.0 system has been in Shenzhen road testing for more than half a year, and has been successfully adapted with The Flying Marvel R, Lincoln MKZ, Geely Geometry A and other models. Zhou Guang believes that this year's focus will not be on expanding his own fleet, since he has come up with a mass-produced L4 system, then this year's focus is on cooperation with OEMs to gradually promote front-loading mass production.
Second, the third path to climb unmanned Mount Everest appears
There is no doubt that in the autonomous driving industry, many people are looking forward to the arrival of the real era of fully unmanned autonomous driving, but how to climb this unmanned Mount Everest, everyone has their own climbing method.
"From the very beginning of our business, we have been thinking about how to achieve a mass-produced L4 autonomous driving system." Zhou Guang told the car something like this.
"In this process, the data is extremely significant for the iteration of autonomous driving." Zhou Guang said.
He said that at present, the industry mainly collects data by building its own fleet and launching a mass-produced L2 system. But there are obvious problems with both approaches.
The cost of self-built fleets is very high and difficult to mass-produce. This means that for many self-driving startups, the road to self-built fleets will not be easy, and the collected data is ultimately inferior to mass production. The launch of the mass-produced L2 system seems to be a delaying strategy, but in fact, because the data structure of the L2-level automatic driving system and the L4-level automatic driving system is different, the data of the mass-produced L2 system actually has no way to promote the iteration of the L4 system.
Therefore, in response to these two problems, Zhou Guang identified the third path with the team at the beginning of his business to climb the unmanned Mount Everest - to create a mass-produced L4 solution, sell it to consumers, and achieve massive data collection.
Seeing this, there may be readers who will wonder, how did Yuanrong Qixing do the automatic driving system of mass production L4? And when will consumers actually use L4 level autonomous driving systems?
▲Yuanrong started the L4 level mass production plan deployment timeline
First of all, in view of the mass production cost and vehicle regulations of the L4 automatic driving system, Yuanrong Qixing adopted a combination of embedded chips and solid-state lidar and cameras on the hardware.
Secondly, Zhou Guang believes that in order to achieve the large-scale rollout of mass production L4 level automatic driving, we should first concentrate on overcoming the most difficult scenarios, which is why Yuanrong Qixing has invested its main fleet in Shenzhen, where almost all of the autonomous driving test areas are in the core area with the most complex and changeable road conditions. Conquered the most difficult scene, and then promoted it later, it is a dimensionality reduction strike.
As for when will we be able to use the mass-produced L4 autonomous driving system? Zhou Guang revealed that at present, Yuanrong's mass-produced L4 products have been discussed with many head car companies, so the earliest time point for consumers to use mass-produced L4 may be 2024.
Third, mass production L4 is not easy to die technology layout in advance is the key
As mentioned earlier, this set of mass-produced L4 products taken out by Yuan Rong Qixing is the first in the industry. In fact, this is precisely because it is not easy to come up with a mass-produced L4 product.
Zhou Guang gave us a few examples, such as in order to reduce costs, the system currently launched by Yuanrong uses solid-state lidar and embedded chips, then there will be two problems - solid-state lidar is not good in performance, after all, it is impossible to achieve a 360-degree detection angle; and the embedded chip due to low computing power, so to achieve high-level automatic driving capabilities, it puts forward higher requirements for the team's algorithm.
So how did Yuan Rong Qixing solve these problems? Zhou Guang told us the answer.
1. Solid-state lidar point cloud data quality is not enough? The algorithm is sufficient.
Since the solid-state lidar point cloud quality is poor compared to the point cloud data of mechanical lidar, the field of view is also relatively small. Therefore, solid-state lidar has not yet been widely adopted by L4 level autonomous driving companies.
Yuanrong Qixing used a number of solid-state lidar to achieve L4 level automatic driving. Zhou Guang introduced that the perception algorithm of Yuanrong Qixing is a pre-fusion algorithm, which can be compatible with different brands and different models of lidar; and the algorithm also has rich experience in the integration of images and lidar features, so it can make the perception module after the use of solid-state lidar accurately identify objects around the vehicle two hundred meters away.
In order to ensure safety, Yuanrong Qixing also designed a pure visual redundant perception system, in the case of partial sensor failure, the system can also adaptively carry out automatic driving according to safety policies.
2. Low computing power of embedded chips? Self-developed inference engine.
In view of the problem of low computing power of embedded chips, Yuan RongQixing gave the solution of the self-developed reasoning engine.
In AI, the inference engine generally includes a scheduler, an executor. The scheduler is responsible for the coordination of the overall computing resources, so that the calculation graph is executed in an optimal manner. An actuator is the execution of an operator. The figure contains the execution of operators, each operator can be generated by an efficient codegen, so that the generated operators and scheduled resource strategies can be the most efficient on the current hardware platform.
Zhou Guang told Che that the so-called reasoning engine could be understood as a compiler that could make the program run faster and more efficiently on it.
At the beginning of 2020, Yuanrong Qixing developed a reasoning engine technology designed for the L4 level autonomous driving deep learning model, so that the automatic driving algorithm can run efficiently and stably on a low-cost, low-power hardware platform, and obtain a reasoning speed of 6 times higher than the mainstream deep learning framework.
Therefore, even if an embedded chip with low computing power is used, thanks to the self-developed inference engine, the algorithm does not need to be optimized for compression, and the system can run smoothly.
Of course, the reasoning engine to achieve is not simple, the construction cycle is long, the technical requirements are high, and the team of Yuanrong Qixing is precisely because it has begun to do this thing at the beginning of the business, so it can take the lead in realizing the application.
In addition, at present, the common difficulties of the entire autonomous driving industry - game and target prediction, Yuanrong Qixing has also made its own innovation.
Previously, Chedong had experienced the L4-level unmanned vehicle started by Yuanrong in Shenzhen, and the most impressive thing was that The unmanned vehicle of Yuanrong was overtaken and jammed during the driving. So the biggest feeling at the time was that the company's algorithm was radical enough.
▲YuanRong set off on the convoy
After hearing the evaluation of the car, Zhou Guang was very confident: "We are still relatively high in the industry in the game and prediction. ”
He said that the game and prediction of Yuanrong start-up are based on deep learning based on mathematical models, so the accuracy rate will be higher and safer, which is why the unmanned vehicles started by Yuanrong will be more "bold and careful" compared to other unmanned vehicles.
Conclusion: Breaking the deadlock of automatic driving, Yuan Rong is one step ahead
Autonomous driving has become hotter and hotter in recent years, but higher-level autonomous driving still seems out of reach – most of the people equipped with L4 level autonomous driving systems are still the Robotaxi fleets of self-driving companies, and the scale of mass production is limited.
On the first day of entrepreneurship, YuanRong Qixing figured out this path to break the game - the realization of L4 level automatic driving must start from mass production. Now, Yuanrong has indeed taken a step ahead - the first to launch a mass-produced L4 level autonomous driving program that can be launched on the road, and is steadily moving forward.
"We've always been determined to do the technology and do what we're good at." Zhou Guang told us this at the end of the interview.