In-depth: When will autonomous driving technology mature? | Intelligent Driving Column

"When will this autonomous driving technology mature?" This is probably the question most often asked by the public when talking about autonomous driving.

Recently, the US government announced that it plans to reduce the number of traffic accident deaths in the country to zero within the year. The background of this plan is that the National Highway Traffic Safety Administration announced that the number of traffic accident deaths in the United States increased by 10.4% in the first half of 2016. Autonomous driving is considered to be the cornerstone of realizing this plan.

"When will this autonomous driving technology mature?" This is probably the question most often asked by the public when talking about autonomous driving.

● ● ●

The road to commercialization of autonomous driving: revolution or improvement?

In March this year, at the SXSW interactive media event held in Austin, Texas, Urmson, the technical director of Google's driverless project, said that "the actual appearance of self-driving cars may be much later than previously predicted, and it may take up to 30 years." In 2009, when Google's driverless project began, Urmson ambitiously expressed the hope that Google's driverless car would be available on the market in 2019.

But Larry Page obviously doesn't have such patience. In a TED show this year, Larry Page said when talking about driverless cars that he had the idea when he was in college and was very excited about the potential of the project. When asked when driverless cars would be realized, he said: "I think it will be very, very soon. I am very keen to launch the product as soon as possible." He further clarified his attitude: "We need revolutionary changes, not improvement changes."

Last year, Google hired Krafcik, a veteran in the automotive industry and former CEO of Hyundai Motor's U.S. branch, as the head of its driverless project. The general interpretation in the industry is that Google will separate this project from X lab and start its commercialization journey.

But where is the way forward? Will Google build its own cars, provide software product licenses, or directly operate like Uber? This is a topic that the industry has repeatedly speculated on. In an interview with Bloomberg Businessweek in July, Krafcik admitted that Google has not yet determined the business model for driverless cars. The outside world's voices are more pointed. Ajay Juneja, CEO of Speak With Me, believes that Google has not formulated a clear commercialization plan; market consulting firm Strategy Analytics bluntly stated that Google needs a partner, a sales team, and a market strategy.

In early August, Urmson left Google in disgrace. The departure of this Carnegie Mellon University robotics expert undoubtedly reflected the huge rift between Google's vision and commercialization in autonomous driving.

Google has stated on many occasions in public that it has no intention of becoming a complete vehicle manufacturer, but is seeking cooperation. However, the reality is that we have seen Tesla's high-profile breakup with Mobileye, a strong ADAS technology supplier, GM's billion-dollar acquisition of autonomous driving startup Cruise Automation, Ford's acquisition of Israeli machine learning and computer vision company SAIPS, and investments in at least three startups related to autonomous driving. BMW, Toyota and others have jointly invested in Silicon Valley startup Nauto. Other car manufacturers are also stepping up their recruitment of talent. Mainstream car manufacturers are launching an arms race for autonomous driving, and they are ready to do it themselves.

It is not easy to cooperate with car manufacturers, especially for the next generation of core technologies. If a company is not prepared to build its own car and is too high-profile in the matter of autonomous driving, it will often backfire and become an obstacle to cooperation. After all, the automobile industry is one of the industries with the highest brand recognition, and car manufacturers do not want to be overshadowed by suppliers.

"Autonomous driving will define the next decade. We predict that autonomous driving will have a huge impact on society, just like the assembly line invented by Ford more than a hundred years ago." Ford CEO Mark Fields said this, which shows the significance of autonomous driving to the automotive industry. Perhaps because of this, Larry Page is so persistent in his desire to achieve fully autonomous driving.

● ● ●

Self-driving trucks are on the rise

But the reality is always unexpected. At the beginning of this year, Anthony Levandowski, a technical genius of Google's driverless car team, and Lior Ron, the former head of Google Maps, left and founded the startup Otto. Only half a year later, mobile travel giant Uber announced in July that it would acquire Otto for a high price of $680 million. Considering that this startup was only established for half a year, it can be concluded that they mainly rely on the technology developed in Google. Now, the commercial value of this technology has been recognized by investors, and the commercialization prospects are very clear: to provide autonomous driving operation services for freight trucks.

Why is there such a huge gap in commercialization of the same technology? The answer needs to be found in the current market situation. Let’s first look at the current situation of the freight market in the United States:

The U.S. trucking industry is worth $700 billion, which is big enough;

In 2015, there were 1.6 million truck drivers in the United States, accounting for 1% of the U.S. working population. The average age of truck drivers is as high as 55 years old, which shows that the industry is not attractive enough to young people and it is difficult to recruit people. There is currently a shortage of 50,000 large truck drivers (according to data from the American Trucking Association), which has become a problem for logistics companies.

Due to the labor shortage, the revenue per mile is expected to rise from the current $0.3 to $0.7.

If autonomous driving is used, for logistics companies, the labor cost savings are undoubted (although it will cause more than 1% of the labor force to lose their jobs), and the expenses of service areas, motels, and restaurants are basically unnecessary, and the cost of auto insurance is greatly reduced; in terms of production efficiency, autonomous driving vehicles can operate 24 hours a day, 7 days a week, and even considering the time consumption of vehicle maintenance and cargo loading and unloading, the driving time of 140 hours per week is more than three times that of human driving. This means that the capital turnover rate is also improved at the same time.

In terms of deployment costs , large trucks are now priced at more than $150,000. Otto's current autonomous driving kit costs about $30,000. Judging from the room for cost reduction, it is very likely to be reduced to $10,000 within five years.

Technically , trucks mainly run on highways, and autonomous driving technology in highway scenarios is much simpler than in urban roads. Sensors can be installed higher from the ground, so they can detect farther. Therefore, autonomous driving in this field can reach the maturity required for commercialization in a short period of time.

Although accurate cost calculation still requires more data, the above data already makes people believe that logistics costs based on autonomous driving will be reduced by more than 2 times, and the return on capital will likely increase by more than ten times.

According to Reuters, Uber will provide freight services starting next year. It has to be said that Uber has a good vision. More companies are starting to follow suit. The six largest truck manufacturers in Europe (including Volvo, Daimler, DAF, Iveco, MAN, and Scania) have formed a fleet of more than 12 driverless trucks for road testing.

Self-driving trucks could profoundly change the logistics industry, giving rise to places like standardized loading and unloading stations, where human drivers would be responsible for handing over goods to self-driving trucks that have just pulled off the highway exit, completing the last 100 kilometers of transportation.

Looking back at Otto's success, it is actually a beautiful example of the combination of technology and market. The research and development of a technology may face an extremely long-term goal at the beginning, but when the technology reaches a certain level, it can generate commercial value in certain scenarios. Therefore, by limiting the use scenarios, the difficulty of technical implementation can be reduced. In the scenario of highway freight, we can even further reduce the technical difficulty, such as using a human-led convoy. In this mode, a human driver drives the lead vehicle, followed by 5 to 10 unmanned vehicles driving at dense intervals. This means a 5-10 times increase in labor efficiency, and its commercial value is evident.

● ● ●

The road to autonomous driving applications

The public's current doubts about autonomous driving are actually based on an assumption that cars can be driven on any open road at any time. This is undoubtedly the ultimate goal, but can you imagine how an autonomous car can obey the instructions of a traffic policeman speaking in a dialect? It is easy to draw a pessimistic conclusion by using the most difficult scenario to evaluate a developing technology. In fact, as we have seen in the example of autonomous trucks, the generation of commercial value is diverse. In the process of achieving the ultimate goal, it is possible to implement autonomous driving applications and generate commercial value by limiting scenarios or functions.

If we explore the commercialization path along the same lines as self-driving trucks, we will find that many industries can achieve it in the short term.

• Warehousing and logistics industry: Leading e-commerce companies such as Amazon and JD.com have deployed AGVs;

  
  

• Autonomous driving vehicles for agriculture: including agricultural machinery that can be used for tilling and harvesting. It is not difficult to move at low speeds on non-roads, and other transport vehicles can be used for transfer;

  
  

• Partially enclosed places: such as resorts, tourist attractions, airports, mining areas, docks, construction sites, etc. Most of the vehicles used in this application are special vehicles, such as excavators, cranes, small electric vehicles, etc.

As technology develops further, more autonomous driving scenarios will be realized:

• Urban public transportation systems: have fixed routes, such as using bus lanes, where autonomous driving can be selectively implemented.

  
  

• Commercial operation vehicles: such as taxis, company shuttle buses, etc.

For private vehicles, the application and popularization of autonomous driving can also be gradually expanded according to different scenarios: for example, first on highways; then parking lots, and finally on open roads. A survey from General Motors shows that in super-large cities, 30% of gasoline is wasted in the process of finding parking spaces, and the parking time in the central city is usually more than 15 minutes. The autonomous driving of parking lots is actually very meaningful, and the driving environment in parking lots is relatively friendly, there are no weather factors, the speed is low, and it is also a closed place, so the difficulty of implementation is obviously lower than that of open roads.

● ● ●

The logic of commercialization

No matter what the application scenario of autonomous driving is, there are always three core principles that need to be met: the technical maturity reaches the requirements of the application scenario; the investment cost is acceptable; the return on investment reaches a breakthrough point: compared with previous manned driving, it must be able to reduce costs or increase revenue, and this commercial benefit can be quantified. In a word: deploying autonomous driving means saving money or making money, otherwise it can only fall into a show scenario.

Therefore, the commercialization path of autonomous driving is bound to be different in different countries, because the same application scenario has a different cost structure. For example, in the taxi industry, the labor cost of taxis in the United States is obviously much higher than that in China. This is one of the reasons why Uber is so aggressive in its investment in autonomous driving technology.

In China, first-tier cities are actively promoting the development of test areas for autonomous driving. The commercial driving force behind this cannot be ignored. More than 20% of the roads in Beijing are occupied by cars, while the actual utilization rate of a private car is usually less than 10%. More than 90% of the time, it is parked, which brings two huge problems: difficulty in parking and congestion (tidal traffic, local road congestion caused by looking for parking spaces).

As for the solution, traditionally, more parking lots and more roads need to be built, which inevitably consumes land. What is the current land price in first-tier cities? The average price in Shanghai is more than 100 million per mu. If we can use autonomous driving to improve vehicle utilization, fewer parking lots can be built. If 100,000 fewer cars are parked, the parking area saved will be at least 1.6 square kilometers, worth 240 billion yuan! The same is true for road resources. If we can use autonomous driving combined with green wave belt traffic to double the road utilization rate, many roads can be replanned, such as reducing the construction of new roads, or demolishing old roads together with communities for overall transformation. The land saving is obvious.

Cost is also not to be ignored. The commercialization of an advanced technology is often accompanied by a gradual decline in cost. Today, most of our cars are equipped with radars, but when radar was first put into practical use, it was on the eve of World War II. In 1936, the British set up the first radar station on the coast of Sophocles. Later, during World War II, radar technology developed rapidly under huge military needs, from ground-based air defense radars to ship-borne radars, and then airborne radars also appeared. After that, radar gradually entered civilian use. Radar was only used in automobiles 20 years ago, and the application of 77GHz millimeter-wave radar in automobiles has only been in recent years. There are certainly technical factors in this, but cost is undoubtedly decisive. After all, cars are mass consumer products, and cost factors cannot be ignored. The same is true for autonomous driving. Before the cost is reduced to a certain level, it is impossible to popularize it in passenger cars.

● ● ●

Technical factors that cannot be ignored

At present, almost all autonomous driving test vehicles are based on the autonomous perception mode, that is, using a variety of different types of sensors to perceive the environment. However, any sensor has limitations, and sensor fusion is not so easy to do. For example, the radar detects a strong echo in front and confirms that it is a large obstacle, but the camera does not see it. At this time, which sensor data should be believed? The real situation may be that there is a can in front, which amplifies the radar echo.

The safety required for autonomous driving is so high that multiple redundant sensing methods are necessary. V2X plays a very critical role here. Unlike cameras or radars, V2X is a precise sensing method. Relying on 802.11p or 5G communication, V2X can accurately sense the situation of surrounding vehicles in a much larger range (300 meters indoors and 1000 meters outdoors), including their location, speed, turn signal status, etc. At the same time, through communication with road infrastructure, it can obtain accurate geographic information in a local area; V2X makes dynamic fleet networking possible, and can achieve green wave band traffic through V2I; V2X will not be affected by weather conditions, all of which are difficult for other sensors to achieve. It can be said that V2X has made a qualitative leap in the reliability of autonomous driving.

However, V2X is a typical technology that relies on standards, which means that there is only one real leader: the government. At present, China's V2X standards are still being formulated, and LTE-V technology has just begun to be studied. There is still a long way to go before it can be put into practical use. This reflects the complexity of government decision-making, which makes it difficult to predict the deployment time of V2X. V2X deployment in local closed places is much easier. A company can deploy V2X at an airport or a certain park, and because it is a closed place, there is no need to consider interoperability or standards, so the efficiency of commercial implementation is very high.

The value of V2X depends on its popularity. If the coverage of V2V on the road is not 100%, its significance will be greatly reduced. V2I requires a lot of investment in infrastructure. At present, the media spotlight is still on car manufacturers in the development of autonomous driving, but if V2X is to achieve full coverage, the role of the government is irreplaceable. Like any large-scale infrastructure investment, the deployment of V2I cannot be planned for full coverage at the beginning, but must start from a local level, because such an investment must consider commercial returns. We can imagine that the government may first deploy V2I on several trunk highways, such as the G2 Beijing-Shanghai Expressway and the Yangtze River Delta Urban Belt Expressway, to improve road utilization and obtain higher commercial returns.

Today, the mainstream technology development direction of the perception and decision-making links of autonomous driving is gradually becoming clear, that is, machine learning based on the combination of deep learning and reinforcement learning, but machine learning needs to be driven by big data to achieve high performance and high reliability. This means that developers need to install autonomous driving equipment on a large number of vehicles first, so that the vehicles can generate the required amount of data in actual operation, resulting in a chicken-and-egg problem: at the beginning, the reliability of autonomous driving is not good, and a large number of equipment cannot be sold; insufficient equipment leads to insufficient data, which will restrict performance improvement. If it is deployed in certain specific applications at the beginning, data can be gradually accumulated, performance can be improved, and preparations can be made for a wider range of autonomous driving applications.

Google's self-driving car has only traveled 3 million kilometers since it began testing seven years ago, while Tesla's Autopolit has accumulated more than 160 million kilometers of mileage in the six months since it was put into use in October 2015. Uber is even more exaggerated. Morgan Stanley once said in a report: "The data collected by Uber in 24 minutes is equivalent to all the data recorded by Google's self-driving cars since their birth."

This is the advantage of accumulating data through mass-produced cars. In the final analysis, data-driven technology must be perfected in practice. The testing scale of a single company is limited and the efficiency is too low. There are countless examples of this in the history of technology. When Apple launched the first generation of iPhone, users found hundreds of bugs just a few weeks after it was launched. This was the work of Steve Jobs, who had high quality requirements.

At the same time, the development of autonomous driving is highly regional. The driving environments in Europe and North America are different, and the driving environment in China is even more different from that in Europe and the United States. For example, drivers frequently change lanes, people and vehicles are mixed, vehicle characteristics vary greatly, and the road system is complex. These local factors often lead to a sharp decrease in the success rate of functions that were originally mature in Europe and the United States, such as changing lanes and crossing intersections. In this sense, localized data processing and the development of autonomous driving decision algorithms cannot be avoided. On this issue, Horizon Robotics has had extensive discussions with quite a few foreign car manufacturers and international Tier 1s, and all parties have highly agreed.

● ● ●

Inspiration from history

In its report "Driving Toward Safety," the U.S. think tank RAND Corporation believes that the safety of autonomous driving requires hundreds of millions to hundreds of billions of miles to verify its reliability, highlighting the complexity of fully autonomous driving on open roads and the huge challenges in testing methods that this brings.

With its outstanding achievements in machine learning, Horizon Robotics has attracted widespread attention in the industry in the field of autonomous driving. At the same time, Horizon Robotics is positioned as a Tier 2 supplier of autonomous driving solutions. It has maintained extensive communication with many car manufacturers and Tier 1 companies, and has carried out in-depth cooperation with many of these partners. Such a relatively upstream positioning gives Horizon Robotics a good perspective to observe and think about the industry, and to formulate corresponding technology roadmaps based more on the real needs of the market.

On the road to the ultimate goal of fully autonomous driving, many intermediate nodes have been identified. We believe that ADAS will continue to evolve, transitioning from the current Level 2 to Level 3, and eventually to Level 4 autonomous driving. In terms of technology research and development, from perception to 3D scene semantic understanding, to environmental situation prediction, path planning, and scene aspects, its application will gradually expand from highways to general roads. At the same time, ADAS will also expand from external perception to perception and understanding of drivers, ensuring the reliability of the transition between autonomous driving and manual driving.

An ambitious enterprise often seeks to achieve success in one fell swoop and commercialize revolutionary technologies in one step. Motorola pioneered the era of mobile communications. At that time, the biggest challenge of mobile communications was the insufficient coverage of ground base stations and the poor reliability of switching between base stations. So when developing a new generation of mobile communication technology, this technology giant decided to build a satellite communication system to completely solve this problem. This is the famous Iridium system. Technically, this is a great idea, but what is unexpected is that the construction speed of ground base stations exceeded expectations, and the reliability gradually improved during the popularization process. However, the Iridium system relying on Motorola alone could not support itself. The huge investment led to high service fees, and the profit return did not meet expectations, and eventually it withdrew from the market.

In a sense, if a system is complex enough, whether it is technical, political or economic, it is difficult to launch a complete and reliable system in one step.

More than 20 years ago, when the iron curtain of the former Soviet Union fell helplessly, Fukuyama made a famous statement: the end of history. He optimistically believed that the liberal democratic system was "the end point of the development of human ideology" and "the last political form of mankind." Today, the chaos in the Middle East proves that the liberal democratic revolution may not necessarily bring a better political ecology, and there seem to be more successful examples of reform. Russia also tried to use shock therapy to rebuild a perfect economic structure, but it failed miserably. On the contrary, China's gradual reform has achieved great success. Amid the clamor of today's autonomous driving, the enlightenment of history may be worth pondering.