Robin Li’s new car, China’s autonomous driving in the past nine years

Original link:

When tech companies and car companies envision a driverless future, they often remove the steering wheel—the most visible sign of a human driver.

Baidu also showcased their self-driving car plans in this way. At the Baidu World Conference on July 21, Baidu released the sixth-generation unmanned vehicle “Apollo RT6”, announcing that it can be mass-produced and hit the road next year.

The RT6 doesn’t look much different than a regular SUV — a big step up from it. The sensors such as the huge lidar installed on the roof of Baidu’s unmanned vehicles have been integrated into the body.

Baidu this car can have no steering wheel. The rear view mirror has naturally been removed as well. As for the driver’s position, it can be used to store luggage or a small desk for office use. We all know that working in the car is an important imagination of the new forces for the future of the Chinese people. Another is to add a screen between the two rear seats.

The Apollo RT6 shown at the launch has a steering wheel. Li Zhenyu, senior vice president of Baidu Group and head of the intelligent driving business group, said that if Chinese law allows it to go public, the option to remove the steering wheel can be provided.

Robin Li's new car, China's autonomous driving in the past nine years

Figure: Apollo RT6 displayed at Baidu World Congress.

Baidu has announced that its production cost — 250,000 yuan — is half cheaper than the previous generation of Baidu’s self-driving cars.

Baidu CEO Li Yanhong said that the cost of Apollo RT6 has dropped, and Baidu can deploy tens of thousands of unmanned vehicles across the country.

The timetable has also been drawn, and today it is very optimistic – it is expected to start a small-scale trial operation in the second half of next year, it is expected to operate in 65 cities in 2025, and cover 100 cities in 2030.

But just nine years ago, when China’s technology and Internet companies first started their self-driving car plans, 2025 wasn’t even radical.

In 2013, Baidu was regarded by many as the most technologically savvy company in China’s Internet industry, and it was the first to recruit people to study unmanned vehicles. In the following nine years, capital and practitioners have slowly combined technological progress with commercialization in the setbacks again and again, turned the direction again and again, and found a way to popularize unmanned vehicles.

The Beginning of It All: Powerful Miracles, Straight to Unmanned Robotaxi

The source of unmanned vehicles is the Unmanned Vehicle Challenge held by the US Defense Advanced Research Projects Agency (DARPA) in 2004. Some of these winners joined Google, secretly developing a self-driving car project (now called Waymo) in 2009, which was revealed by the US media the following year.

Baidu began to develop unmanned vehicles in 2013. Using the data interface provided by BMW, it developed its own software and installed Velodyne’s lidar. It took more than two years to transform a BMW 3 Series GT into a driverless car. , drove from Baidu Building to Beijing Fifth Ring Road, and then returned the same way. During this period, I followed the car to slow down, change lanes, and overtake many times, with a maximum speed of 100 kilometers per hour.

Robin Li's new car, China's autonomous driving in the past nine years

Figure: Baidu’s first-generation unmanned vehicle based on the BMW 3 Series sedan.

There was no problem in the test, and Baidu took the opportunity to set up an autonomous driving division. Like Google, Baidu also has optimistic expectations: in a few years, it will develop an unmanned vehicle (generally called L4 in the industry) that does not require human intervention, and then take passengers on the street (called Robotaxi), subverting taxis and online booking. The car industry is reshaping the way people travel.

After the two search giants made progress in the research and development of autonomous driving, a group of autonomous driving entrepreneurs emerged. Both the two big companies and the people who left them to start their own businesses are almost convinced that there should be no ambiguity about who controls the car, the human and the algorithm.

There is a widely circulated story in the industry that Google launched AutoPilot (L2), a semi-autonomous driving system, in 2013 and gave it to several employees on long commutes to test. Those workers are expected to keep their eyes on the road, stay alert and respond to accidents when the semi-autonomous system is working. But they quickly forgot the road conditions and used computers, slept, and put on makeup in the car at nearly 90 kilometers per hour. Concerned about safety, Google stopped the test after only a few weeks, and hasn’t resumed since.

Semi-autonomous driving looks like the car is already driving itself, making it difficult for people to stay focused. Therefore, many researchers believe that driverless technology should wait until the real automatic driving is perfected before handing it over to users.

Capital is convinced of this. In 2018, Morgan Stanley valued the spin-off Waymo at $175 billion, more than twice that of Uber, and approaching the market value of Toyota, the world’s largest automaker at the time. According to Yiou statistics, China’s self-driving companies raised 16.2 billion yuan this year, more than 20 times that of two years ago. At that time, Baidu released its unmanned car in Las Vegas, USA, and many reporters could not find their seats and sat on the ground to watch.

Optimists believe that the mass production of unmanned vehicles will become a reality in the next year, and even if the process is tortuous, it will run all over the street in five years. But they are ultimately overestimating the development speed of the self-driving car industry.

The end point is far away, and the problem of unmanned driving is not only a technical problem, but also a cost problem

Unmanned vehicles rely on sensors such as lidar, millimeter-wave radar, ultrasonic radar, and cameras to collect information about the surrounding environment of the vehicle (pedestrians, other vehicles or obstacles, etc.), and then pre-store it in the system in advance to record road information (lane lines, sidewalks, With the help of high-precision maps such as traffic lights, etc.), it can determine its own position, and then input this data into the artificial intelligence model, so that it can quickly make judgments with the help of chips with powerful computing power, plan the driving route, and then control the vehicle’s actions.

It is not the most difficult thing to develop an unmanned car that can run on a simple and closed road. Baidu started from scratch and built a team of 20 people, and it took two years to make it. What is even more difficult is how unmanned vehicles can respond quickly and safely in various extreme environments.

For example, in heavy rain and foggy weather, the accuracy of the sensor is damaged. How can the vehicle perceive the surrounding environment information and ensure the safe driving of the vehicle? When a vehicle is driving at a high speed, a pedestrian suddenly rushes in front of the vehicle, how to respond quickly?

These extreme problems are being solved. Li Feifei, a professor at Stanford University, has published a paper with team members, trying to use a computer vision system to judge the intention of roadside pedestrians.

But the development of unmanned vehicles is not only a technical issue, but also a cost issue.

The iteration of unmanned vehicle technology requires a large amount of road test data, so that the system can learn to deal with various extreme situations that cannot be envisaged in advance, so as to minimize the risk.

Early self-driving cars cost more than $100,000 for hardware alone. Collecting data from just one vehicle is not enough. With the slow progress of research, the industry generally believes that at least tens of billions of kilometers of road test data are needed to develop large-scale usable unmanned vehicles. This means that 10,000 cars equipped with lidars and computing chips will drive 10 hours a day at a speed of 40 to 50 kilometers per hour for 10 consecutive years.

It is naturally not cost-effective to let tens of thousands of vehicles run empty, and these companies also need to hire a special person to sit in the driver’s seat to monitor the vehicle and intervene in the event of an accident – the salary cost of the safety officer for each vehicle will end up being more expensive than the car.

In order to cut costs as much as possible, unmanned vehicle companies have introduced simulation technology to construct virtual road environments in computers to train unmanned vehicle systems. At the same time, Baidu proposed automatic “parameter tuning” technology in 2018. The unmanned vehicle model can automatically optimize the system based on the feedback of road test data, instead of relying on people to grope and adjust by feeling.

However, the road test is still necessary and the cost is still high. Coupled with the killing of people by Uber’s unmanned car, the industry’s optimism turned to pessimism, and the patience of capital was gradually worn away. In 2019, Waymo’s valuation was cut by 40% to $105 billion, and the amount of financing in China’s autonomous driving field fell by a third year-on-year. In 2020, when Waymo brought in investors outside of Google, it was rumored to be valued at only $30 billion, only one-sixth of its peak value.

Large companies with resources have deployed multiple directions ahead of time, trying to generate revenue on the long road. In 2017, Waymo tested the Robotaxi for free, and started testing autonomous trucks the same year.

Chinese self-driving car companies have not fallen far behind.

Li Zhenyu, head of Baidu’s intelligent driving business group, said in an interview with “LatePost” that in March 2017, Li Yanhong internally suggested that Baidu should take the lead in building an open platform for unmanned vehicles and introduce more companies to innovate separately.

A month later, Baidu launched the Apollo program, attracting universities or institutions with research capabilities and companies with application scenarios to develop various types of unmanned vehicles together.

In the following years, Baidu continued to develop Robotaxi, sent them to test roads for autonomous driving in various cities, and cooperated with other companies to start unmanned minibuses, unmanned trucks, unmanned mine trucks/port vehicles, and unmanned vehicles. Express vehicles, unmanned sanitation vehicles, etc. The choices of unmanned car startups are similar, and most of them will choose one or two directions to focus on.

Technology companies hope that these attempts will be closer to the current needs of users and gain a portion of the revenue to support subsequent research and development. Compared with unmanned vehicles driving on urban roads, the driving environment of the above-mentioned vehicles is relatively simple and controllable, with lower technical requirements, making it easier to turn unmanned vehicles into reality. But there are also many competitors in these fields, and it may not be able to make much money in the short term.

Participating in building cars has become the only way for driverless cars

Tesla caught up. With millions of Teslas driving on the road every day, and the owners of the cars serving as test safety officers for free, the car company can collect more data faster and more than the self-driving company without much extra cost.

In 2013, Tesla began to study Autopilot. At that time, Musk also said that it is an “assisted driving system”, which is called L2-level automatic driving in the industry, and only assists humans to complete some driving tasks under certain conditions.

After the unmanned car on Google’s route attracted attention, Tesla’s Autopilot iterated to version 2.0, and Musk’s propaganda was more aggressive, calling the assisted driving system, which still has serious flaws, “full self-driving.”

Tesla’s assisted driving system relies more on the data obtained by the camera. It insists on not using lidar, and the cost is lower. It has embarked on a “mass production route” that is sold with the car. Both the Model 3 and the Model 3 are equipped with hardware that supports assisted driving systems.

In two years, more than 300,000 Tesla vehicles have been on the road, collecting data for Tesla. And users also need to pay for Tesla’s immature driver-assist features — up to 64,000 yuan ($1,000) per car (the current sticker price).

As of November 2018, Tesla has obtained more than 1.6 billion kilometers of Autopilot driving data. Although the quality is not as good as the data collected by the road test, the mileage alone has far surpassed any company specializing in the development of autonomous driving.

Not-so-mature assisted driving is open to ordinary car owners, and the capabilities of this system are deliberately promoted. There are indeed many situations that unmanned car companies try to avoid. According to Tesla Deaths website statistics, 15 people have encountered when driving Autopilot. Died in a car accident.

The accidents did not affect the increase in Tesla testers. In 2021, Tesla will sell 930,000 vehicles and generate $3.8 billion in revenue from its Autopilot software system.

The new domestic car-making companies Xiaopeng, Ideal, Weilai, etc., also use self-driving technology as a selling point for their cars, and even take all of Tesla’s propaganda. It was not until August last year that a Weilai car accident caused the industry to reflect.

But the industry is reflecting more on marketing, and the trend of cramming autonomous systems into cars continues. At the same time as the Weilai incident, Xiaomi, which had just announced that it had built a car for less than half a year, announced the acquisition of an autonomous driving company.

Car manufacturers start with assisted driving and then iterate to the progressive route of unmanned driving, which effectively solves the problem that traps the development of unmanned vehicle technology companies – the hardware and labor costs of road test data collection can be done by consumers. .

Many automakers are already putting lidars—sensors that were once only needed for driverless systems—in mass-production cars they sell to consumers.

Some unmanned vehicle practitioners are beginning to feel that the biggest competitor in the industry may no longer be peers or car-hailing companies, but car manufacturers.

Baidu released Apollo Lite, an unmanned driving system that does not rely on lidar, since 2019. Like Tesla, it uses the information obtained by the camera to judge road conditions and control the car. The “Pilot Assisted Driving System” developed by Baidu based on this technology will be loaded into some new cars of WM Motor, BYD, Dongfeng Lantu and other automakers.

Robin Li's new car, China's autonomous driving in the past nine years

Figure: The Apollo Lite system analyzes 10 cameras in parallel with a data volume of 200 frames per second.

In May of this year, WeRide announced that it had invested in Bosch, a first-tier supplier in the automotive industry, to provide software development and technical support, and develop low-level (L2-L3) assisted driving technology to be sold to automakers.

More aggressive companies choose to end up building cars. Baidu has come up with two plans. One is to form a joint venture with Geely in March last year to form Jidu Automobile, recruit engineers from scratch to design vehicles, and plan to produce them independently, equipped with Baidu’s automatic driving function and in-vehicle system, and released a concept car last month, a mass-produced car It is expected to be released next year.

The second is the simultaneous preparation of the Apollo RT 6 project. Baidu said that building a car is inevitable for the development of unmanned technology, and refitting a car can no longer improve the capability of unmanned vehicles. If you want to ensure more safety redundancy, you can only go further upstream and enter the automotive design process.

And driverless technology is also improving the driving experience of vehicles. Although artificial intelligence automatically picks you up from get off work is still considered a science fiction story, but some driverless technologies have been dismantled and integrated into people’s lives.

Most of the new cars released in the past two years have the “auto-parking” function, and some cars can already park the car in the parking space automatically without human intervention. Baidu’s autonomous parking system (AVP) has entered a period of high volume for multiple customers, and has been used in mass-produced vehicles such as Weimar W6, GAC Egypt Safety Series, and Great Wall Haval.

The hardest thing is always trust

In 2004, at DARPA’s first Unmanned Vehicle Challenge, the most powerful unmanned vehicle could only travel 12 kilometers in open desert terrain. Today’s unmanned vehicles have entered urban roads, and unmanned vehicles have appeared on the streets of Phoenix in the United States, Beijing Yizhuang and other places (the safety officer is in the co-pilot and in the cloud). Anyone can download the application and click on the screen to shout Get a driverless car.

Robin Li's new car, China's autonomous driving in the past nine years

Picture: The strongest player in the first DARPA unmanned vehicle challenge, a Hummer modified by the Carnegie Mellon University team, drove 12 kilometers in the desert.

Robin Li's new car, China's autonomous driving in the past nine years

Figure: Baidu’s fifth-generation mass-produced self-driving car, Apollo Moon, is unmanned while driving, with a safety officer sitting in the co-pilot.

The cost of producing a self-driving car continues to drop. When Google and Baidu first started to develop unmanned vehicles, the price of only one core sensor 64-line lidar was 800,000 yuan. Now domestic lidar companies have cut the price to 10,000 yuan. Baidu said that the control of the cost of Apollo RT 6 at 250,000 is the result of the development of the entire industry chain. They are discussing OEM cooperation with multiple car manufacturers, and it has reached the advanced stage.

At present, there are nearly 1,000 unmanned vehicles that can carry passengers on China’s roads, of which Baidu has nearly 600. It has picked up millions of passengers for free, and has started toll travel services in many cities across the country.

Since Google secretly built unmanned vehicles, technology companies, car manufacturers, and startups around the world have burned about $50 billion to develop unmanned vehicles. The most difficult thing for unmanned vehicles to finally hit the road is not technology or cost.

Xu Baoqiang, general manager of Baidu’s Automotive Robotics Department, said that in addition to designing seven redundancy for Apollo RT 6, including power supply, communication, braking, steering, and sensors, to ensure that there will be no accidents as much as possible, another important task is to study the country and the country one by one. Local traffic regulations are guaranteed to meet all requirements.

“It’s more difficult (to put people) into orbit than to launch a rocket,” former Waymo CEO John Krafcik said in an interview last year. Human spaceflight only needs to be safe once, and self-driving cars need to keep passengers safe over and over again. Can’t go wrong.

Krafcik was born in the traditional auto industry and has been in charge of Waymo for more than 5 years. He has experienced the industry from optimism to pessimism. He is one of the calmest practitioners in the unmanned vehicle industry. drving license”.

People naturally trust their peers more. Even if people make low-level mistakes while driving from time to time, or even drink and drive, 1.3 million people are killed in traffic accidents every year, and no one proposes to ban humans from driving on the road. But even a fatal accident with artificial intelligence driving the road will cause repeated discussions.

Unmanned driving must be far safer than ordinary people before it can be accepted by the public and legally permitted.

While progress has not been as fast as initially hoped, few in the industry today doubt that driverless cars are finally here. Like all new technologies, autonomous driving is coming more slowly and faster than people think.

This article is reprinted from:
This site is for inclusion only, and the copyright belongs to the original author.

Leave a Comment

Your email address will not be published.