[Pacific Auto Network Industry Channel] Not long ago, a photo of an unmanned express delivery car caught in the undried cement pavement spread all over the Internet. Unmanned driving technology, which is given high hopes by people, an unmanned vehicle that integrates the most advanced technology of the moment, fell into embarrassment under the most simple scene of "cement is not dry, forbidden to pass"...
When "unmanned driving" encounters "cement is not dry", is there a solution?
The incident occurred in the Jinming Campus of Henan University, and the protagonist of the incident was a "little wild donkey" unmanned express delivery car developed by Ali Damo Academy, and the little wild donkey has been serving in rookie stations across the country for more than a year. In 2021", Ali announced that it has deployed thousands of small wild donkey unmanned express delivery vehicles in more than 200 colleges and universities in more than 70 cities across the country, which is currently one of the most numerous and wide-ranging unmanned vehicles put into use in China.
The little wild donkey unmanned delivery vehicle was launched at the Yunqi Conference on September 17, 2020, with a body length, width and height of 2100/900/1445mm (including lidar height), and the perception is equipped with 1-4 lidar, 6 cameras, millimeter wave radar, inertial guidance and other sensors in front and rear, with an emergency response speed of 7 times that of humans, with L4 level automatic driving capabilities, with the goal of solving the delivery problem of the last 3 kilometers of delivery.
Such a perception hardware configuration, in the unmanned vehicles currently in use, belongs to the near "full match", and all aspects of the attributes have been pulled full. At the same time, unlike the common passenger car automatic driving, the average speed of the little wild donkey is only set to 15km/h, and the maximum speed is only 20km/h, which also reduces the potential risk factor of its unmanned driving. The little wild donkey put into operation can deliver nearly 600 couriers per day, and relying on the huge volume of rookie, its actual operating mileage and accumulated data volume are also among the top in the global autonomous driving related research enterprises.
However, the more "pretty" the parameters of the little wild donkey, the more awkward the performance in this incident. Why is such a hardcore cutting-edge technology so frustrated in such a simple scene?
Autonomous driving relies on cameras, lidar, millimeter wave radar and other sensors to make decisions about the perception of the surrounding environment, and each sensor has its own advantages and disadvantages, complementary performance, and can identify potential obstacles in the direction of travel quite effectively. But whether it is a camera or a radar, it is impossible to make an effective judgment on whether the cement on the smooth road has hardened.
In this incident, we can see from another angle of the photo that the construction party set up a warning sign in the middle of the road that reads "Construction in front is prohibited from passing", and there is no effective warning in the direction of the small wild donkey's travel, no warning sign, no pile tube, etc. The other three sides have slender warning bands, and the warning belts in the direction of travel are tiled on the curb side. Since there is no video of the incident, we cannot judge whether the warning belt has fallen when the little wild donkey has passed, or whether the little wild donkey has failed to effectively identify and crashed into the warning belt. It can only be said that this slender warning band has the possibility of not being effectively identified by the automatic driving sensor. Since there is no conclusion, I will not expand too much here.
It is worth noting that "cement undried" is not only a problem that automatic driving will encounter, we often see kittens, puppies, birds and other animals on the cement floor to leave a clear mark, a string of particularly eye-catching. Even for normal humans, it's not uncommon to see stories of rushing into unhardened concrete floors with inadequate warnings. Human intelligence is not yet completely accurate, let alone a toddler's unmanned vehicle?
At the same time, the cameras (vision), lidar, millimeter wave radar and other perception sensors relied on for automatic driving do not have the ability to judge whether the cement has hardened, and under the condition of insufficient warning, "cement is not dry" for automatic driving, it can be described as insoluble.
The "Long Tail Effect" of Autonomous Driving
The unmanned express car falling into the cement floor itself is indeed an embarrassing event that people can't help but laugh at, but this matter has aroused people's attention not only to the event itself, but also to worry about the hot "automatic driving" technology today.
In just a few years, the field of artificial intelligence has made amazing progress with the blessing of technologies such as deep learning algorithms, big data, and more powerful semiconductor hardware. With this, people are developing rapidly in the field of autonomous driving. From Google's heavy investment in the development of driverless cars, to Tesla's popularity around the world, to today's self-driving / high-end assisted driving has become an indispensable research direction for almost all car companies. We have also witnessed the emergence of autonomous driving related technologies from the laboratory to our side step by step.
In the classification of automatic driving, L2 level and below belongs to assisted driving, the automatic driving ability of the system is weak, and the driver's expectation of the ability of the system is also low, which can ensure the safety of use. L3 level and above belongs to automatic driving, in the design operation area (ODD), the driver's expectation of the system gradually increases, and the system also has enough safe automatic driving capabilities, and the expectation and ability are equal to ensure the safety of use.
However, under the current conditions, whether it is the L2+ assisted driving that has been widely equipped with passenger cars, or the so-called L4 level unmanned delivery vehicles that are truly unmanned, they are still in the stage where the technology is not yet fully mature. It has to be said that the current automatic driving related technology has been able to handle most of the daily scenarios, but the risk of automatic driving lies in the unpredictable "long tail".
"The Long Tail" was originally used in 2004 by Chris Anderson, editor-in-chief of Wired, to describe the business and economic models of sites such as Amazon, Netflix, and Real.com/Rhapsody. Refers to those products or services that were not valued in the past, but the total revenue accumulated exceeds that of mainstream products due to the huge total amount.
Extending the "long tail effect" to the field of automatic driving is to say that the potential risks in most of the common head scenarios in daily life have been solved in the daily training of autonomous driving, but those unexpected scenarios that are not taken seriously are extremely rare, but there are many kinds, so the cumulative total has also posed a great threat to the safety of automatic driving.
The scenarios that self-driving cars need to face can be classified into 4 categories: Category 1, known safety scenarios; Category 2, known unsafe scenarios; Category 3, unknown unsafe scenarios; And Category 4, unknown security scenarios. Of all the scenarios, the most difficult is the type 3 scenario, which is the unknown unsafe scenario. The long tail effect generally occurs in category 3 unknown unsafe scenarios.
Of course, engineers are also constantly trying to reduce, or even completely eliminate, the third type of scenario, the most common way is to use the supercomputer, the daily extremely difficult to encounter, but may appear dangerous scenes, in the supercomputer to carry out a large number of scene simulations, the results of the training into the automatic driving algorithm, so as to avoid the actual encounter with similar scenarios when the danger.
Taking Tesla as an example, at the 2021 "AI DAY", Tesla introduced some rare scenarios, such as an extreme case of a wind and snow rolled up by a truck in front of it obscuring the vehicle in front. In order to solve this rare but unusually dangerous event when it occurs, Tesla used a supercomputer to simulate more similar scenarios and multi-neural networks for extensive training.
Tesla's daily solution to this kind of "long tail" scenario is far from unique. Musk said they simulate all sorts of rare cases that can be imagined, including "elk walking leisurely on city roads" and even "flying saucer falls", which is a strange danger that cannot happen at all.
But the scary part of the "long tail effect" is that it is so large and unpredictable that no amount of effort by engineers can exhaust these potential risks. For example, the simple scene of "cement not drying" makes the little wild donkey fall into an embarrassing situation.
What's even more frightening is that "unscrated cement" is now a Class 2 "known unsafe scenario", but the algorithm does not have an effective means to change it into a dangerous scenario that can be avoided.
C-V2X may be an effective solution to the problem
What is C-V2X? Let's first understand what V2X is
V2X, or "Vehicle to Everything", is an IoT communication technology that connects a vehicle to other things, where V stands for the vehicle and X represents the object of information that interacts with the vehicle. The information modes of V2X interaction may include: V2V (car and car), V2I (car and road), V2P (car and person), V2H (car and home), V2N (car and network), V2C (car and cloud) and so on.
In order to achieve V2X, there must be a unified communication standard, the most popular of which is the DSRC and C-V2X two standards. The United States has been promoting the DSRC standard for a long time, and China has fully implemented the C-V2X standard.
On December 12, 2019, the Federal Communications Commission (FCC) unanimously voted on a proposal that would reallocate most of the spectrum in the 5.9GHz band and dedicate that spectrum to unlicensed spectrum technology and C-V2X technology. It should be noted that for a long time, 75MHz in the 5.9GHz band has been designated for DSRC. As a result, C-V2X has gradually become the most widely accepted V2X communication standard in the world, and it is expected to become the only standard in the world.
With C-V2X, vehicles can communicate with subgrade equipment and other vehicles on the surrounding roads, and many problems that are difficult to solve by vision and radar sensors have become very simple under V2X technology. For example, the "ghost probe" problem, which is difficult to effectively solve whether it is automatic driving or human driving, is actually just a simple and limited field of vision. Through the surrounding multi-angle street lights, telephone poles, and other vehicles on the road, cameras, radar can achieve almost no dead angle perception of an area, through C-V2X technology to achieve information sharing, "ghost probe" has become a false proposition that is no longer valid.
The "uncoiled cement" scenario we mentioned earlier only requires a simple V2X communication device to draw a precise construction area "electronic fence", and autonomous vehicles can easily avoid this area.
In fact, not only these simple scenarios, the Internet of Everything based on the C-V2X standard can play a greater role in the "smart city", and can be efficiently controlled for urban traffic diversion, some regional hazard warnings, etc. For example, if there is a flood or a road subsidence in a certain section of the road, it can be sent to every vehicle that may go to this area through the smart city cloud, so that it can avoid danger in advance.
V2X is so good, why hasn't it been popularized so far? Because it's hard!
The key to V2X is collaboration, a vehicle, a road equipped with related communication equipment can not play its due value. So V2X also fell into the "chicken or egg first" problem. At present, the support of countries in the world is different, and earlier car companies tried to take the lead in achieving the popularization of V2V through several major groups, but it is obvious that only the V2V promoted by some car companies cannot play its due ability in the face of huge car ownership.
At present, the mainland is vigorously developing subgrade C-V2X equipment, through government leadership, so that car companies and consumers have the enthusiasm to promote and purchase vehicles with C-V2X capabilities. However, the popularity of C-V2X did not happen overnight. The road is obstructive and long, and C-V2X is not yet a realistic solution to solve the "dry cement".
epilogue
Science and technology is the key force to promote human progress, and the sophisticated can enter the reality but it is inevitable to encounter some simple embarrassments. Cutting-edge high-tech equipment, occasionally need a roll of ordinary tape to stick a handful... The most advanced driverless, encountering "cement undried" will really fall into embarrassment. But technology never happens overnight, and it is with the help of these embarrassing mistakes that it can continue to grow.
Is there a solution to the "unmanned driving" encountering "cement not drying"? Of course, as long as enough eye-catching and effective prompts are set in advance after construction, such as a few more pile tubes, the current unmanned vehicle can also be easily identified. If the warning is insufficient, the next person to fall into it may be a passerby who is not vigilant enough. (Text: Pacific Auto Network Guo Rui)