“There will be 1 million driverless taxis on the road in 2020.” In April 2019, Tesla CEO Elon Musk, who has always been “big mouth”, declared this on Twitter.
Interestingly, Kai-Fu Lee commented under this tweet that if there are really 1 million Tesla self-driving cars on the road by then, “I will eat them all.”
Looking back now, the various beautiful fantasies and fanatical optimism surrounding autonomous driving are gradually fading. Everyone knows that in the gradual development process of the automotive industry, autonomous vehicles will eventually arrive in the near or far future. But at present, when the four wheels of this unmanned vehicle will actually land on a large scale, it is still There is no clear timetable.
In the post-human era when constantly trying to break through the limits of the physical body, emerging technologies are not only “writing” humans themselves, they also bring unprecedented changes to the daily transportation scene.
In the past more than a century, the products of the automobile industry have been designed around the core of technology driven by engines and using fossil fuels as energy sources. The strong entry of “outsiders” smart electric vehicles has been rapidly deployed in the industrial chain links of complete vehicles, parts and energy storage systems, and has become a trend that has driven tremendous changes in the global automotive industry in recent years.
The advancement of science and technology has continuously carried out more and more comprehensive arms on automobiles. Statistics show that almost 10% of the world’s total economic growth is related to travel and transportation. In 2017, Intel released a report stating that by 2050, the global autonomous vehicle market is expected to reach 7 trillion U.S. dollars.
Carrying the imagination of “technology creates a better life”, there is such a huge commercial potential in front of us, and the industry is full of confidence inside and outside. At the end of 2016, the two major technology giants, Apple and Google, successively abandoned their long-prepared car-building plans, and instead invested heavily in the research and development of autonomous driving and intelligent interconnected software businesses.
The global car market at that time looked like the eve of the mass production era of autonomous vehicles: new energy car companies represented by Tesla were the most active in integrating software and hardware of autonomous driving technology; traditional car companies such as BMW , Volkswagen, using its own profound advantages on the manufacturing side, united with technology companies to accelerate the development and technical reserves of smart cars; and core players of artificial intelligence such as Nvidia and Horizon-chip vendors, are also in the “arms race” of autonomous driving AI chips. Start a fierce competition.
Once autonomous vehicles are proven to be feasible, a large number of driverless taxis on the road become a reality, which will have a disruptive impact on the inherent business model of producing and selling cars. After all, under the premise of technical guarantee, most users will not choose to be tied up with a seat belt and sit behind the steering wheel.
On January 21, 2002, San Francisco, USA, C ruise launched a self-driving car called Or ig in
What’s more, people are more willing to believe that artificial intelligence will always be stable, there is no possibility of “drunk driving”, and will never commit human repeated crimes compared to human drivers who are tired, emotionally fluctuating or even “road rage.” mistake.
On March 19, 18, investigators were inspecting the Uber self-driving test vehicle in the accident
The craze spread quickly, and the auto industry giants fell into collective anxiety. In the transition from a mere car manufacturer to a mobile travel service provider, traditional car companies have not been much slower than the “new car-building forces”. In March 2016, General Motors acquired Cruise, a start-up company engaged in the development of autonomous driving technology, for a high price of US$1 billion.
Numerous “Cruises” popped up overnight, and “wide spreading of coins” became the norm. Traditional car companies are holding money everywhere to find people to invest. There is even news that there are start-ups related to autonomous driving that can get a considerable amount of financing based on the resumes of their founders.
Increasingly blind optimism spread, and the practitioners did not wake up until the world’s first fatal accident of autonomous driving occurred.
/ The autopilot system’s performance in operating details such as stopping brakes and predicting obstacles in advance is bad. /
On March 19, 2018, the self-driving test vehicle of the ride-hailing company Uber was involved in a car accident in Arizona, USA, killing a woman who was crossing the road with her bicycle. At the time of the incident, the test vehicle that caused the accident was in an automatic driving state, and there was a safety driver in the vehicle.
After waiting anxiously for two months, the National Transportation Safety Board announced the results of the accident investigation. According to its disclosure, 6 seconds before the collision, Uber’s test vehicle, which was in the process of autonomous driving, had detected the victim, but identified it as an unidentified object, a car, and a bicycle.
Later reports revealed that when the incident occurred, Uber took the initiative to turn off the emergency brake function on the test vehicle to reduce the abruptness of system control and driving behavior. Uber’s administrators argued that this move is also a manifestation of safety measures: If an unmanned car brakes or turns sharply at the wrong time, it will seriously affect other vehicles on the same road.
But this still exposes an aspect of autonomous driving technology that urgently needs to be optimized. In recent years, autonomous driving technology is far from being able to identify obstacles on the road with a zero error rate, let alone perfectly simulating the smooth driving behavior of humans.
To try a commercial driverless taxi, a safety officer is still required in the car. This is definitely not a high-tech scene that people imagine that it will be recruited and the car is in order. After handing the accelerator and brake to artificial intelligence to handle, the developers found that compared to human drivers who have been familiar with almost ignoring their existence for so many years, the autopilot system is in operational details such as stopping and predicting obstacles in advance. The performance is bad, resulting in unspeakably comfortable ride experience for passengers. The inability to create an advanced modern travel environment is also contrary to the original intention of inventing autonomous driving.
If the actual technical problems are not solved, even if the test ride experience is good, it will only bring more doubts and distrust to the outside world.
The former star start-up travel company Roadstar completed the A round of US$128 million in financing in 2018, setting a record for a single financing for a Chinese autonomous driving company. According to reports, a person in charge of a capital institution involved in the A round of financing decided to invest after only 40 minutes of trial experience in Silicon Valley.
However, Roadstar’s self-driving car test ride experience is excellent, but it shows that its self-driving system used for road testing, the driving strategy set is more radical than the judgment in daily driving-the world’s first fatal accident caused by automatic driving, part of the reason It is Uber that shuts off the vehicle’s emergency braking system in order to avoid excessive unnecessary braking during driving.
People have gradually realized that autonomous driving technology cannot be commercialized in the short term. The capital market is getting colder, and most start-up companies aiming at driverless taxis have difficulty financing, and many have already chosen to exit the market.
Classic misjudgment: is it a car or a cloud
The landing of self-driving cars is hindered, and the time for large-scale launches will continue to be postponed because the technical difficulty exceeds previous expectations. You must know that a reliable autopilot system must have smart chips, algorithms, and low-latency communications as technical support. Simply put, it must be “intelligent” and “connected.”
If you want an autonomously controlled car to ride in turbulent traffic safely and at high speed like in a sci-fi movie, starting from the current technology, it is necessary to install intelligent hardware such as radar and camera on the car body to check the road conditions, obstacles and location. Wait for information to perceive, and then perform image recognition and analysis through the core algorithm, select the execution plan, and finally perform vehicle control and operation actions such as deceleration, braking or steering.
Cruise, which is favored by giants, has high hopes in the industry, and in July 2019 announced the postponement of the unmanned taxi business plan originally scheduled to debut that year.
On May 19, 2015, at the site of the German unmanned driving technology press conference, the staff demonstrated the automatic parking assistance application
Road testing is the key to building a deep learning model for unmanned driving. The longer the road test mileage, the richer the driving scenarios experienced by the test vehicle, which means more driving experience data for the system to learn. However, when Cruise’s unmanned vehicles were tested in 2018, the technical bottleneck that could not accurately determine the state of roadside objects is still difficult to solve.
When passing a row of motorcycles or bicycles parked on the side of the road, Cruise’s test vehicle would hesitate, not knowing whether it should stop during the journey; sometimes it could not identify pedestrians; sometimes it mistakenly thought that there was a bicycle ahead and took the act of stopping. .
The technical bottleneck encountered by Cruise in developing its autonomous driving system is not accidental. At the moment, the biggest limitation hindering the development of autonomous vehicles is still perception.
Nowadays, the common equipment of vehicles’ perception hardware includes lidar, millimeter wave radar and camera. They are extremely susceptible to interference from external factors such as weather and light, not to mention that the automatic driving system also needs to comprehensively distinguish the information collected by different sensors.
On the early morning of June 1, 2020, a truck rolled over on the highway in Chiayi, Taiwan, China. While the truck driver was waiting for rescue on the side of the road, a Tesla Model 3 that was coming from behind slammed into the truck box. The front of the car until the A-pillar was almost completely submerged in the truck.
On July 3, 2020, the autonomous driving demonstration zone in the North District of Zhongguancun Science City, Beijing
Prior to this, there have been several similar traffic accidents in Tesla vehicles with Autopilot driving assistance turned on. A very intriguing common feature of this type of accident is that the damaged surface of the crashed stationary vehicle is all white, and the bottom is almost suspended. In the eyes of Tesla’s sensors, this looks like an unobstructed road ahead under the blue sky and white clouds.
Faced with a large area of white car body, coupled with strong light reflection during the day, it is difficult for the camera to extract the effective feature points of the obstruction ahead. When the millimeter-wave radar cannot achieve fine recognition, the system can easily The white truck standing in the middle of the road ahead was judged as barrier-free.
This is also true. Even in the background of clear weather and high visibility, although the camera can work normally, the system has sufficient computing power, and it itself has gone through a long period of machine learning, but the recognition failure still cannot be eliminated. The probability event. This is an extreme situation that cannot be avoided by all autonomous driving systems based on vision solutions. Although its possibility will be continuously or even infinitely reduced during the system’s machine learning process, as long as it happens once, it is difficult for people to trust the machine itself with complete peace of mind.
Endless extreme scenarios
In layman’s terms, as long as the machine has not learned before, there is a probability that an accident will occur. But in reality, various extreme scenarios can hardly be exhausted. In the industry, it is generally believed that 8 billion miles of driving tests are required to ensure that unmanned driving systems are safer than humans. If you use a fleet of 100 vehicles to calculate, you need to operate 24 hours a day for 100 years to achieve the goal. This is an impossible task in the near future.
/ As long as the machine has not learned before, there is a probability that an accident will occur. /
However, optimists believe that there will be a certain stage in the future where 5G technology and unmanned vehicles will interact, and the gradual progress path led by auto companies will also become the main driving force for the real implementation of autonomous driving technology. The logic is that autonomous driving cannot be put into use until everything is ready and perfect. Only gradual progress can promote the iterative development of technology.
But there is more than one technical difficulty that needs to be overcome. The perception technology of an autonomous driving system requires the ability to process multiple information in addition to high requirements for recognition rate. At the very least, its reaction speed must not be slower than that of a human driver.
In addition, there are still ethical problems to be solved in the ethical context, such as the “tram problem” of whether a few people should be sacrificed to save more people, waiting for the auto industry and policy makers to answer. Autopilot seems not far from real life, but before landing, there is a thick stack of test papers waiting to fill in the blanks. This is also the reason why human drivers are still the safest decision-makers for vehicles.