Written by | Zhao Lingwei
Edit | Chen Hao
Produced | Automotive Sankei
Recently, Li Xiang (founder, chairman and CEO of Ideal Automobile) and Xia Yiping (CEO of Jidu Automobile) debated the number and location of lidar on social platforms.
As a new category, lidar currently has a variety of different technical routes in the industry. For car companies, choosing a suitable program for themselves and letting the products land as soon as possible is the king.
The entire self-driving industry (except Musk) is waiting for lidar to mature, because only cameras, millimeter-wave radar assisted driving has indeed had a lot of accidents before this.
Not long ago, on the Yueyang National Highway in Hunan Province, a Xiaopeng P7 that turned on auxiliary driving crashed into a truck that flipped over on the road without slowing down at all - the same auxiliary driving, the same rollover truck, which is the same as the two accidents of Tesla at home and abroad in the past two years.
Xiaopeng Automobile's response was that it was initially judged that the owner did not maintain the observation of the environment in front of the vehicle and take over the vehicle in time during the use of ACC+LCC.
Coincidentally, this year's plan of Xiaopeng Automobile to push the city NGP (Intelligent Navigation Assisted Driving) has been on the string. Not only Xiaopeng, but also the HI version of the jihu and Huawei models will also launch similar high-end assisted driving functions.
As a result, when the new generation of high-end assisted driving began to hit the road, people's judgment of it was constantly pulled between doubt and trust.
Standing in the first year of high-end assisted driving, are car companies really ready?
NO.1
["Can't see clearly"]
Vision + millimeter wave radar has bugs, and lidar to be verified
This is not the first time that the Xiaopeng P7 has been "planted" on assisted driving.
In September last year, a Xiaopeng P7 rear-ended with the front board with the NGP turned on. At that time, the technology blogger @ Feng Kai analyzed that "the two redundant detection systems of the perception module did not find the presence of obstacles".
At that time, assisted driving perception systems were usually composed of cameras + millimeter wave radar, and cameras needed to use a large number of samples to train target detection. However, small probability situations such as flatbed trucks in front of them and trucks that overturn on the road are very small in the actual training data, and it may not be recognized by the camera when the accident occurs.
The ability of the camera can only continue to explore, the current ideal L9 equipped with the camera has reached 8 million pixels, but in the range of more than 500 meters for object recognition or stay on large objects such as vehicles, for pedestrians, cone barrels and so on still stay at about 200 meters. The real human eye, on the other hand, can recognize the cone barrel 400 meters away and react.
However, the pixels of the camera are too high, and the car chip will fall into the black hole of computing power, and the camera, chip computing power and algorithm will form a situation of rolling themselves.
It is difficult to solve various corner cases (long-tail scenes) only by relying on camera recognition, even if it is the "algorithm ceiling" Tesla, its pure visual autopilot scheme has also had many accidents of all sizes, such as recognizing the overturned truck as a white cloud, causing irreparable tragedy.
The partner's millimeter-wave radar also has many problems in perception - first of all, there is a relatively large omission for static and relatively static obstacles, and secondly, the detection of millimeter-wave radar is too sensitive, there will be many false inspections, and alarms are everywhere.
Some industry experts said that when the two are combined, vision is generally used as the main sensor and radar as the auxiliary sensor. So, it may be that the radar detects an obstacle, but the vision does not recognize it, the decision-making system chooses to believe the vision. At this point, the driver's takeover is very important.
In short, it can only be said that the sample collection and training are insufficient. In the product brochures of Xiaopeng, Weilai, Tesla and other car companies, a large number of driving assistance systems with low recognition or cannot be enabled are displayed.
Corner cases are currently a challenge for all automated/assisted driving systems. Both cameras and millimeter-wave radar have flaws in the recognition of irregularly shaped and stationary objects, which require lidar to compensate for it — a small thing that is considered the glory badge of L3 and even L4 level autonomous driving.
However, according to Zhu Xichan, a professor at the Automotive College of Tongji University, no lidar can really achieve mass production at present, and even if it can be mass-produced, the cost is quite high. "The driverless cars purchased by the intelligent networked car demonstration area and the high-resolution lidar perceived by the road end are not committed to the 'warranty period' by many enterprises." At present, the 16-line mechanical scanning lidar and the all-solid-state, hybrid solid-state lidar used on the mass production model can still ensure reliability. However, these mass-produced lidars do not have high angular resolution enough, resulting in a detection distance that is not far enough. ”。
As a lidar company, Su Shuping, general manager of Innoviz China, recently confirmed this statement - lidar has not yet been a fully mature stage into large-scale mass production.
She said in an interview, "I think at least 500,000 units a year, lidar can call large-scale mass production on the car, so this time period I think it is still from 2024 to 2025, or even after."
In terms of effect, Zhu Xichan said in an interview with Automotive Sankei, "At present, the angular resolution of all lidar cannot support the detection of road debris above 300 meters, so it cannot support HWP --- 130km/h automatic driving on the highway." ECE R157 ALKS is a highway TJP function, the current lidar products are able to support 0-60km/h speed range of automatic driving perception and response, efforts, some companies' products can achieve the highway maximum speed of 80km/ h automatic driving. However, for HWP, the user's expectation is automatic driving at a maximum speed of 130km/h. ”。
If you change the scene to a more everyday urban road, the requirements for hardware capabilities may not be so high, but the pressure comes to the software side.
NO.2
["Not accurate"]
The AI library is difficult to complete, and the algorithm is inaccurate
The chief scientist of the Argo Lab Artificial Intelligence Driverless Research Center at Carnegie Mellon University attributed the delay in autonomous driving to two points: the underlying technology and real application scenarios.
From the current situation, the development of underlying technology can already cope with most traffic scenarios, but the real road conditions are still unpredictable, and the development of autonomous driving technology is trapped in this 5% of the corner case.
According to expert analysis, the algorithm needs a large number of road test scene experiments to be continuously tested and optimized, and the more scene data the algorithm's decision-making will be more accurate, but if there is no corresponding scene in the library, the algorithm may have no way to do.
But it is almost impossible to enter all scenarios into the autonomous driving system, which requires the ability of machines to think, that is, breakthrough development in the field of algorithms and digital.
Zhu Xichan said, "The high-definition camera of automatic driving can only distinguish between 'people' as such 'wireframes', even if it is recognized as a person, but it does not deal with the direction of the person, the state of arms and legs, it is difficult to judge the person's movement intention in the next 5 seconds, and the human driver is still very strong in judging the pedestrian's movement intention according to the attitude and even the eyes of the pedestrian, combined with experience."
That said, AI has a hard time dealing with surprises.
Fortunately, in the face of "unknown unsafe scenarios", in early April, the five ministries and commissions issued the rules of the Notice on the Trial Implementation of the Automotive Safety Sandbox Supervision System, and launched the pilot work of automotive safety sandbox supervision.
As a flexible regulatory system for technological innovation, sandbox supervision can provide an enterprise with a test platform and test cycle, and on the basis of not violating the principled access standards and the bottom line of supervision, encourage enterprises to voluntarily carry out further testing when they do not fully grasp the product risks, so as to prevent product application risks to the greatest extent.
Continuously improving the database through sandbox supervision is the current general direction, but the "accident database" can never be 100% perfect, and the system of car companies can only continuously reduce risks.
From the perspective of car companies, the solution to this problem is very clear - find people. Algorithms are one of the core capabilities of autonomous driving, and the core of algorithms is professional talents.
"No company (autopilot) is doing it from scratch, Baidu autopilot before and after four groups of people have been poached, Xiaopeng and Tesla lawsuits are happy, including Cruise people are constantly job-hopping, the whole world is like this."
Zhu Xichan told Auto Industry and Economics that the talents of various companies have repeatedly jumped sideways, so that their technical routes have gradually become similar, and their capabilities have become more and more likely to the same level.
For example, the starting point of Xiaopeng NGP is the entry of Wu Xinzhou, vice president of research and development, Weilai gave up the closed Mobileye at the end of last year, and dug up the algorithm team of R&D director Ren Shaoqing to do the automatic driving camera, and only then did it really start full-stack self-research.
Recently, there has been a new wave of turmoil in the autonomous driving-related talents of Weilai, Ideal, Xiaopeng and Tesla, which may be related to capital.
For now, Zhu Xichan believes that the greater significance of full-stack self-research lies in attracting the attention of capital. After all, the most expensive thing in the 21st century is talent - some car companies that attach importance to autonomous driving have allocated hundreds of millions of yuan of "digging people" budgets.
The world's top talents have not yet understood the automatic driving game for more than a decade, and the difficulty is evident.
The defects of the algorithm and the problems caused by the unsafe machinery and equipment also make it difficult for high-level automatic driving to land.
NO.3
["Ouyang Feng's Three Questions"]
"What's my name?" "Where am I?" "What am I going to do?"
Not only Xiaopeng, Tesla, Weilai and others claim that the manual of L2+ class assisted driving models mentions that the automatic assisted driving system may not recognize stationary objects. At the same time, the driver must be ready to take over the vehicle at all times when assisting driving.
The official account of @Weipai, which is certified by Zhihu, wrote in the reply, "In 2021, there should be only one fuel vehicle that can be fully L3 level autonomous driving, WEY." Mocha has done completely hands-free (automatic driving) on the road that can be used for testing in the state regulations, so it cannot be opened out, which is easy to affect the judgment of users, and does not dare to elaborate, for fear of violating the law and misleading users. ”
As an academic expert in the industry, Zhu Xichan told Auto Industry and Economics that the L2+ model he tried had a very good experience of "automatic driving" in the L2+ NP state.
For high-level assisted driving, Zhu Xichan summed it up as "Ouyang Feng's Three Questions"——
"What is my name": autonomous driving or assisted driving;
"Where am I": L2 or L3;
"What am I going to do": Can I get the driver off? Can drivers be allowed to watch WeChat?
The most dangerous is also here.
While claiming their own superb level, high-level assisted driving functions repeatedly emphasize that the driver should be ready to take over the responsible vehicle at any time after the accident, so they can only use a series of ambiguous words such as "assisted driving", "high-level assisted driving", "automatic assisted driving" and so on to frantically imply users, but until now, it has not been able to clearly answer the soul torture of "who is it".
The greater risk of autonomous driving at present is that human trust is greater than system capabilities, which is also the "uncanny valley" statement proposed by Zhu Xizan.
This year, for the L3 level of automatic (assisted) driving models or will get a formal "name" after the examination. It is understood that the certification regulations on L3-level automatic driving will most likely be issued within the year.
Although the promotion of L2+ by car companies is pushing users to the bottom of the "uncanny valley", when the L3 with full beard and full tail really opens on the city road, we will experience safe travel that has never been experienced before.
Next time, I hope that smart cars can help owners see the big truck that overturns on the road, and bring people into a more comprehensive era of machine-assisted intelligent driving - in the scene that meets the conditions of automatic driving, the driving right can be safely handed over to the automatic driving system, and the driver can be accurately "called" back to pay attention to driving when there is a risk.
This article guides experts: