Tesla, once the
leader of the "autonomous driving" concept, is now facing the
embarrassment of losing this propaganda tool.
In mid-September, Tesla became the defendant in a class-action
lawsuit in San Francisco. The reason for the lawsuit is that Tesla has misled
the public by falsely promoting the Autopilot and Full Self-Driving (Full
Self-Driving) functions.
Plaintiff Briggs Matzko said in the lawsuit that Tesla’s
move was to get people “excited” about its cars to attract investment, boost
sales, avoid bankruptcy and drive up the stock price.
Last month, Model 3 owner Toledo also sued Tesla in a
California court in the United States, saying that his Tesla would suddenly brake
when there were no obstacles at all, also known as "ghost brakes." He
thought it was a "terrible nightmare".
Screenshot of NHTSA Lawsuit
More than individual lawsuits, on September 2, the
California Senate passed a bill to ban car companies from using the word
"FSD" (Fully Autonomous Driving). The bill is still pending final
signature by Governor Newsom.
California Senator Gonzalez said the bill is very clear
on Tesla. She believes that Tesla has long warned of the limitations of its
assisted driving system in small print, but users clearly did not accept it.
"People in California think FSD is fully automatic,
but it's not at all."
Gonzalez's concerns represent a longstanding controversy in
the self-driving car space. Driven and "educated" by car companies
such as Tesla, people have the illusion that self-driving cars seem to be
within reach. The future of unmanned vehicles is approaching.
But the accident and the resulting crisis of public
opinion have long accompanied Teslas. As long as an accident occurs, people's
suspicion and anxiety about technology, algorithms, and artificial intelligence
will increase.
Or those questions of general concern: Is autonomous
driving safe? What is keeping human beings safe?
How far are we from Tesla's "full
self-driving"?
accident cause
In 2022, when autopilot manufacturers are eager to build
momentum for new products, several traffic accidents have caused consumers to
doubt the technology again.
On July 6, a crash caused by assisted driving technology
occurred in Alachua County, Florida. The car reportedly swerved into a parked
tractor while on the road using Tesla's Autopilot mode. The driver and a
67-year-old passenger were killed.
Many accidents have two things in common: the driver is
mostly distracted or has no hands on the steering wheel. Most of the cars with
the assisted driving function turned on collided with stationary objects.
For example, the U.S. National Highway Traffic Safety
Administration (NHTSA) recently announced that it is investigating a total of
16 accidents involving 15 injuries and 1 death caused by Tesla Autopilot in the
past year.
The agency reported that most of Tesla's 16 crashes
occurred at night. "Autopilot software has certain recognition loopholes,
ignoring stationary objects such as warning lights, ice cream cones and
illuminated arrow boards."
Compared with software loopholes, industry insiders
believe that the excessive publicity of car companies brings consumers
expectations that do not match reality, which is a more dangerous situation.
Hong Zexin, a senior practitioner in the market direction
of a leading car company, told Yancai that the above-mentioned accident was
"actually caused by many manufacturers bringing wrong perceptions to
consumers."
He believes that some car companies tend to announce
demos (samples) with cool autonomous driving functions at new product launches,
making people think that autonomous driving can be achieved. However, in mass
production, low-profile assisted driving systems are often introduced out of
consideration of price and consumer acceptance.
Hong Zexin introduced that the mainstream assisted
driving systems in the industry mainly use two types of sensor fusion
perception: camera and millimeter-wave radar.
Millimeter-wave radar mainly detects obstacles through
radio wave reflections, but it is "too sensitive and often produces false
alarms." To keep the system running smoothly, the algorithm usually
ignores the radar echoes when the road surface is not moving, making it
difficult to identify stationary objects. This is also the reason why many
intelligent driving vehicles are currently "downtime" in the face of
stationary objects.
In contrast, Tesla adopts a more maverick pure visual
perception scheme, that is, only relying on the camera as image acquisition,
and relying on the on-board SoC chip for real-time computing.
Hong Zexin said that the perception scheme that relies on
cameras also has shortcomings. Visual perception is susceptible to
environmental disturbances such as sunlight and alternating light and dark.
Also, “the camera is like a black box, and once it goes
wrong, it’s hard for humans to explain what went wrong,” he said.
The perception system with obvious shortcomings makes the
assisted driving vehicle extremely limited. A reporter from Yanjing Finance and
Economics found that, including Tesla and other high-end assisted driving
systems, the current configuration is difficult to identify relatively
stationary obstacles such as ice cream cones. In the user manual, these car
companies will mark the scenes with limited technology, reminding consumers to
pay attention and observe the road conditions at any time.
Hong Zexin said: "Nowadays, the public does not have
a clear understanding of assisted driving below the L3 level. Assisted driving
is more about saving people's lives when their lives are in danger. For
example, if a person is sleepy, it will turn the steering wheel a little to
prevent the car from deviating from the lane. . . But people now treat it as a
foolproof, no-fault feature."
Repeated car accidents remind people of the distance
between imagination and reality. Li Xiang, founder and CEO of Ideal Motors, once
posted on his Moments: “We call on the media and industry organizations to
unify the Chinese nomenclature standards for autonomous driving. It is
recommended to unify the names: L2 = assisted driving; L3 = automatic assisted
driving; L4 = automatic driving; L5 = driverless."
L2 includes adaptive cruise, lane departure warning, AEB
automatic emergency braking, etc., and is currently widely used in
mid-to-high-end products such as Tesla and "Weixiaoli". From L2 up,
L3 to L5 can be called autonomous driving. At the L5 level, "fully
autonomous driving", that is, no driver intervention is required in any
scenario.
So far, in the Chinese market, no car company claims to
have mass-produced L3-class vehicles, and none of the cars on the market are
self-driving cars. But Tesla, which took the lead in naming its assisted
driving software FSD (Full Self-Driving), has greatly raised expectations for
autonomous driving technology.
In the early days of the industry, the standard is still
unclear
In any case, good market expectations have pushed car
companies and technology companies to compete for the L4 autonomous driving
track to a certain extent. According to some media statistics, in the first
seven months of 2022, autonomous driving companies in China have received more
than 60 financings in total, which is a rare field that frequently receives
financing.
Compared with the L3-level model that still requires a
limited takeover by the driver, various competitors are mainly striving to be
the L4-level vehicle, that is, an automatic driving mode that does not require
a driver in a specific design operating domain.
In the market, two development paths are being generated.
One category is represented by Tesla. Accumulate data and
technology from mass-produced L2 vehicles to progressively develop L4
autonomous driving. Another category is Google's Waymo and Baidu, which are
determined to "leapfrog" autonomous driving in one step.
Zou Dicong, vice president of autonomous driving
technology of Guizhou Hankes Company, told Yancai that the components of
autonomous driving technology can be divided into three modules: perception,
decision-making and control. The difficulty now lies in the perception
prediction and decision-making module of the system.
"Through sensing equipment such as lidar, we can
clearly see and detect various objects on the road." Zou Dicong said. But
how to make decisions after seeing the road surface and how to make the vehicle
run in a smarter way is the challenge facing the current autonomous driving
technology.
"For example, there is a car on the road. When I
want to change lanes to the left, it is unknown how it will react and whether
it will give way. How can artificial intelligence predict other people's
actions and reactions from massive data? It is difficult to make decisions
similar to the human brain," said Zou Dicong.
If the above capabilities are on unobstructed roads, it
is not difficult for the algorithm to plan a safe and efficient path. However,
when faced with complex traffic flow and scene road conditions, artificial
intelligence lacks an understanding of the overall road conditions and cannot
predict the future behavior of surrounding obstacles. Therefore, problems such as
planning trajectory jumps and collisions often occur.
Hong Zexin told Yancai that decision-making and planning
are the core capabilities of autonomous driving technology for L4 companies.
But the industry is in the early stages of development, and there is no unified
methodology.
In short, "the standard for judging a good drive,
there is currently no".
In addition to decision-making and planning, some
industry insiders believe that another difficulty in autonomous driving
technology lies in coordinating and managing complex systems.
Liu Xuan, vice president of Shenzhen Yuanrong Qixing
Technology Co., Ltd., told Yancai: "Autonomous driving is a very
complicated project. It organically integrates various modules to achieve a
low-cost level without manual intervention, and then integrates all elements.
Very difficult."
Different decision-making and overall planning
capabilities determine whether L4-level autonomous driving will give people an
experience like a novice or an "old driver". Hong Zexin said that
self-driving companies are faced with a dilemma whether to adopt conservative
or radical algorithms.
"If self-driving vehicles are too cautious, driving
on urban roads will indeed hinder traffic and give users a bad experience."
Despite the lack of fixed standards within the industry,
both Liu Xuan and Hong Zexin said that safety is still the primary
consideration in the field of autonomous driving at this stage.
"At present, there is one thing in common, the speed
of the car is slow. No company has dared to increase the speed radically."
Hong Zexin said.
True FSD is still far away
Technical challenges, in the eyes of industry insiders,
are not difficulties that autonomous driving cannot overcome.
For autonomous driving companies, what is more urgent is
how to roll out products to the society.
In 2018, Elon Musk once said that Tesla could soon
achieve fully autonomous driving. "Autopilot, the autonomous driving
system, will soon be able to support traffic such as traffic lights, stops and
roundabouts with full autonomous driving capabilities."
Obviously, in 2022, the fully autonomous driving technology
touted by Musk will not change much from 2018.
Yang Shengbing, a professor of new energy and intelligent
vehicles at Wuhan University of Technology, told Yanjing Finance that to
realize the full life cycle layout of R&D, design, manufacturing, operation
and maintenance of L4 and L5 vehicles, infrastructure, personnel investment, regulations
and public awareness are required. of maturity.
These all take time.
Long cycle means blindly burning money. Now, many
L4-level autonomous driving companies have been unable to hold back and are
scrambling to come up with low-cost solutions to achieve mass production.
On July 21, Baidu, the leading company in the industry,
released a prototype of the sixth-generation unmanned taxi (Robotaxi), claiming
that the manufacturing cost of the vehicle was 250,000 yuan. This is completely
different from people's impressions in the past, where hundreds of thousands of
high-cost L4 cars were equipped with lidars at every turn.
The sixth generation of unmanned taxis (Robotaxi)
Low-cost solutions have been put forward one after another
in companies that are committed to making pre-installation mass production
solutions for car companies. In December 2021, Shenzhen Yuanrong Qixing
released a front-loaded L4-level autonomous driving solution with a cost of
less than US$10,000. In June 2022, Qingzhou Zhihang, headquartered in Beijing,
launched a new generation of L4 mass-produced vehicle autonomous driving
solutions, reducing the cost to 10,000 yuan.
Liu Xuan told Yan Finance that the low-cost front-loading
mass production strategy has been established since the early days of the
company's establishment. The reason is that compared with self-built fleets,
the solution is sold to different car companies to achieve mass production, and
the speed and efficiency of data collection are higher.
"Autonomous driving needs to face a lot of problems,
especially the long tail scenario (corner case). To solve these problems, we
can only accumulate a large amount of data and continuously train, iterate and
improve the algorithm." Liu Xuan explained.
In the field of artificial intelligence, the quantity and
quality of data is the key. The deep learning of artificial intelligence is to
obtain complex mathematical equations in a space with thousands of dimensions,
through massive data training, and then achieve the set goals.
However, extreme road conditions (i.e., long-tail
scenarios) that are rare on daily roads, due to the lack of data by various
companies, have become a pain point that is difficult for autonomous driving to
solve for a while.
Liu
Tao, the co-CEO of Zhiji Automobile, once explained that during the driving
process of the car, 90% of the journey will only encounter tens of thousands of
normal road conditions, which only require hundreds of engineers to overcome,
and the capabilities of various car companies converge.
"But the real challenge is that there are more than 1 million long-tail, low-probability extreme road conditions that are very difficult to cover."
Judging from international
experience, no self-driving company has the confidence to deal with various
extreme scenarios. This is also the reason why companies are scrambling to mass-produce
and carry out road tests with all their might.
Sacha
Arnoud, director of software engineering at Waymo, has said that from his
experience, the top 90% of technical work is only 10% of the total work time.
And to complete the last 10% of the work, it takes 10 times more effort.
Therefore, Liu Xuan judged that, based on the technical
characteristics of the long tail scene, even after mass production of L4
vehicles, laws and regulations may not allow them to be called L4, and they
will not allow unmanned vehicles immediately.
"L4-level autonomous driving technology still needs
to pass the test of time and data." Liu Xuan said. He therefore believes:
"In the future, we will go through the stage of human-vehicle co-driving
before transitioning to true unmanned driving."
This judgment is similar to Liu Tao. He once said that
machines still need to learn iteratively, so for a long time in the future, we
will still be in the stage of human-vehicle driving.
.jpeg)
0 Comments