How far has Baidu Apollo gone on the road to autonomous driving?

On October 10, Baidu’s self-driving taxi service Apollo?GO officially opened in Beijing. Friends in Beijing can place orders through Baidu Maps or the Apollo?GO APP at the Haidian and Yizhuang sites. Take a free test ride.

This new gadget has attracted the attention of many people. Among them are automotive media, industry practitioners, college students, geeks, and technology bloggers. They are queuing up at several sites at the test drive site. The number of daily call orders exceeded 2,600, and some people even waited for several hours without getting a ride.

Presumably, everyone wants to know how far Baidu’s so-called “autonomous driving” has developed, and want to see with their own eyes whether the fantasy RoboTaxi has become a reality. However, although Baidu claims that Apollo? GO is fully open, there are still many restrictions on this "comprehensive" -

First of all, the operation period is limited, from Monday to Sunday from 10:00-16:00. It avoids the morning and evening peak hours and the driving environment at night; secondly, the vehicle has preset routes and can only get on and off at 15 pre-planned stops (similar to a shuttle bus); thirdly, the vehicle speed will be limited to less than 60km/h; every The trolley can only accommodate 1-2 passengers in the back seat, and the age of the passengers must be between 18 and 60 years old. The most important thing is that each vehicle is equipped with a safety officer to ensure that it takes over the driving of the vehicle at any time.

Autonomous driving, which is said to surpass L3 and catch up to L4, has given so many restrictions. How far can Baidu’s Apollo? GO achieve? How far has Baidu gone in the field of autonomous driving?

To what extent has Baidu’s autonomous driving developed?

Baidu’s autonomous driving project started as early as 2013. By 2015, it had invested a total of 20 billion yuan. In 2017, it broke into the public eye with a ticket from Beijing’s Fifth Ring Road. Before the road test was launched in Beijing, its driverless taxis had already conducted trial operations in Changsha and Cangzhou, picking up more than 100,000 passengers.

Just a month ago, on September 15, a driverless taxi without a safety officer carried Li Zhenyu, vice president of Baidu Group and general manager of the intelligent driving business group, and a CCTV reporter in Beijing. Driving in Shougang Park.

On the same day, Baidu CEO Robin Li predicted the timetable for the commercialization of autonomous driving at the "Baidu World 2020 Conference". "Within 5 years, autonomous vehicles will enter the full commercial stage." "Intelligent transportation systems can improve traffic efficiency by 15%-30%." "Within five years, China's first-tier cities will no longer need to restrict traffic, and within 10 years, traffic congestion problems will be solved."

Behind the ambitions, there is indeed sufficient R&D strength and experience to support it.

In the 2019 autonomous driving road test report card released by the California DMV on February 27 this year, the MPI value (Miles Per Intervention, average mileage without human intervention) ranked Among the top five companies, three are from China (Baidu, Auto?

In addition, in the competitiveness ranking list compiled by Navigant Research for autonomous driving companies, Baidu also entered the "Leader" rating for the first time, which is at the same level as Waymo and Cruise.

It can be said that Baidu’s research and development capabilities in the field of autonomous driving are not only among the best in China, but also have been recognized by the global industry.

How is Apollo technology implemented?

In terms of hardware, an Apollo test vehicle includes an industrial computer IPC (including a dedicated GPU), a GPS positioning system, an IMU inertial system, a CAN bus interface card, and a large-capacity hard drive (database). In addition to the hardware of these infrastructures, sensors also include top rotating Lidar (128 lines), forward camera (binocular), side camera (monocular), front/rear Lidar (16 lines), left and right Lidar (16 lines) line), front/rear millimeter wave radar, ultrasonic close-in radar, etc., they are basically armed to the teeth.

At the software level, Apollo customized the Linus kernel (4.4.32), but did not list specific application software, databases, etc. For map positioning, Apollo uses Novotel’s GPS and IMU combined positioning system, which can be fused using the Kalman filtering mechanism to provide positioning and attitude information with sufficient frequency and accuracy. This is also one of the best positioning technologies currently.

At the algorithm level, Apollo joined the DeepDrive deep learning autonomous driving industry alliance and released Apollo's autonomous driving data set ApolloScape. Through massive, high-quality real data, iterative updates are made for autonomous driving algorithm development and testing. Baidu said that the amount of data in ApolloScape is more than 10 times that of similar domestic data sets.

The combination of software and hardware constitutes Apollo’s autonomous driving system.

Among them, the core of the entire system is the industrial computer IPC. This configuration specifically uses the Neousys? 6108GC industrial computer, matched with Nvidia's GTX1080 graphics card. The IPC receives data from various sensors through USB and Ethernet interfaces. After processing, it is then Connecting to the CAN card through the PCI interface ultimately drives vehicle movement.

Because Baidu is an Internet company, it does not have the technology and capabilities to produce vehicles. Therefore, the chosen carrier for this assisted driving system is Lincoln's MKZ.

Why this model? Here is a brief explanation.

First of all, the electrification structure of Lincoln MKZ is relatively complete, with control-by-wire throttle, brake-by-wire, and steering-by-wire systems; secondly, there is currently a Dataspeed company whose main product is ADAS tool kits, and Lincoln MKZ The CAN bus protocol is cracked and encapsulated into ADAS Kit for developers; secondly, for the purpose of expanding travel service providers, Ford is willing to provide relevant interfaces to autonomous driving companies, which is one of the few options available.

Based on the above elements, converting Lincoln MKZ into an autonomous driving test vehicle is currently the most convenient and affordable option. In addition to Apollo, autonomous driving companies such as Nvidia and Pony AI also use Lincoln MKZ.

In general, with these technologies, Apollo can already achieve the L3 level of structured roads, or L3 autonomous driving in specific scenarios and limited scenarios. Judging from the feedback from experiencers in Beijing, Apollo can smoothly complete turns, lane changes and U-turns without the need for manual operation by safety personnel in most cases; if there is a vehicle ahead that is too slow, Apollo will also change lanes and overtake; road When a pedestrian suddenly appears in the middle, Apollo can actively slow down and wait for the pedestrian to pass before driving.

However, there are also shortcomings. Because the AI's driving "habits" are not humane enough, its operations when turning and avoiding pedestrians are very stiff. It often makes forced turns or brakes suddenly, which is prone to bumps. Some passengers said that Slight motion sickness.

It can be said that Baidu Apollo still has technical strength, but there is still a lot of room for improvement. So what are the levels of the top car companies in the field of autonomous driving?

The progress of other competitors

Currently, companies developing autonomous driving technology are mainly divided into three categories. The first category is traditional car companies or auto parts manufacturers, including General Motors and BMW. , Audi, Continental, Bosch, etc.; the second category is Internet technology giants, including Google, Baidu, Didi, Alibaba, etc., of which Didi is also a travel platform and has dual identities in the development and use of autonomous driving; the third category It is an entrepreneurial technology company, including Auto?X, Pony.ai, etc.

Here we will give two typical examples, one is Tesla, the representative of the car company, and the other is Waymo, which is owned by Google. The two examples also happen to be representatives of the two technical routes of lidar and visual recognition.

Let’s talk about Waymo first. Waymo is a self-driving car project launched by Google in January 2009. It later became independent from Google in December 2016 and became a subsidiary of Alphabet.

Just at the beginning of this year, Waymo announced that its self-driving road tests exceeded 20 million miles; in March, Waymo announced that it had received its first round of external investment, ***2.25 billion US dollars, and its pre-investment valuation reached 10.5 billion billion, far exceeding Volkswagen, Daimler, General Motors and other car companies.

Compared with Baidu’s Apollo GO, Waymo’s RoboTaxi was earlier. The Waymo One service was launched in Phoenix, USA at the end of 2018. This was also the beginning of global Robotaxi commercialization attempts. In October this year, Waymo announced that it would enter fully driverless driving (without safety personnel) in Phoenix and provide paid services to the public. This marked the first step in the full realization of Robotaxi technology and business.

Because it does not produce cars, Waymo’s biggest advantage is focused on algorithms. It is a solution based mainly on lidar, adhering to the "traffic as a service" business model, focusing on lidar and high-definition maps. The advantage is that it can quickly build system prototypes, not only has low dependence on big data, but also is easier to detect. , extract and segment data, and the security is more guaranteed. However, the disadvantages are that the cost of the sensor is high, the scalability is weak, and the commercialization speed is slow.

In contrast, Tesla's Autopilot represents a "car as a product" service model, gradually transitioning from assisted driving to fully autonomous driving, emphasizing cameras, visual recognition, and light maps. Because it requires visual learning, it relies more on data. For this reason, Tesla also has its own "unique secret" - shadow mode (claimed to have 10 billion miles of measured data), which has the advantages of lower cost and greater scalability. Stronger and faster commercialization; the disadvantage is that safety control is not as good as lidar.

Technically speaking, Tesla is currently the only car company that develops everything from software algorithms to hardware architecture in-house, and its software and hardware technology has always been at the forefront of the industry.

When Nikkei BP dismantled the Model 3, it concluded that Tesla has been six years ahead of its competitors in electronic architecture; Thomas Ulbrich, a member of the Volkswagen board of directors, also admitted that Tesla has made great progress in electric vehicles and development In terms of software, it is 10 years ahead of Volkswagen.

The main advantages of Tesla Autopilot lie in neural networks, massive data and control algorithms. Although it is difficult to distance itself from Waymo or Cruise in theory and technology, it has deep experience in combining algorithms with vehicle control. It is worth noting that Tesla’s Autopilot strategy for entering the market is more radical. Just this month, Tesla announced that it will push the fully autonomous driving version FSD? Beta to a small number of target users to achieve near-L4 autonomous driving.

From the actual test experience, FSD? Beta can achieve "zero intervention driving" most of the time. It can identify traffic lights and prohibition signs on the roadside at intersections, and can automatically select according to intersection markings and navigation. Lane. When passing through complex intersections such as roundabouts, FSD? Beta can also autonomously obey the intersection yield rules and autonomously avoid pedestrians and non-motorized vehicles on the roadside.

At this stage, narrow roads without any markings are a nightmare for all other driving assistance systems, but the camera-led FSD?Beta is still capable of detecting road boundaries and traffic paths. All vehicles parked on both sides can be detected. Even in parking lots with vehicles on both sides, FSD?Beta can still identify almost all traffic participants. During the night test, FSD? Beta's recognition ability remained roughly the same as during the day, and it was still quite accurate.

How far are we from true driverless driving?

In general, Baidu Apollo is already in a leading position in China and is also in the first echelon internationally, but it is still some distance away from industry leaders Waymo, Cruise and Tesla’s Autopilot. It is also Baidu’s goal after Apollo.

So how far are we from driverless cars?

This depends on when the difficulties ahead will be resolved. For example, a technical corner case (Corner Case). The data collected by the vehicle through radar or cameras is uploaded for the machine to learn. However, in actual driving, some road conditions that are beyond the machine's experience range will inevitably occur. These are corner cases (such as the example of Taiwan's Model 3 crashing into a truck). A research report from the 100 Electric Vehicles Association points out that today’s driverless technology can handle 90% of regular road conditions, but the remaining 10% corner cases have a huge impact and require 90% of the time to solve.

For example, the issue of legal responsibility. The subject of responsibility is a crucial concept in any law. However, autonomous driving technology has blurred the division of this concept. If a self-driving vehicle is involved in a car accident, does the driver have to be held responsible? Is it a technology provider? Or does it depend on the brand of the vehicle? These are currently unresolved problems (such as the example of pedestrian death caused by Uber in the United States).

For example, the formulation of rights of way and road rules. Do self-driving vehicles enjoy the same rights of way as human-driving vehicles and are subject to unified management? Do they drive in the same lane and apply the same traffic rules?

For example, changes in the nature of the product. Autonomous driving will greatly increase the utilization rate of vehicles, thereby reducing the stock of vehicles on the entire road, because people will no longer need the ownership rights of vehicles, but only need to have the right to use vehicles. Will the public accept such a change in nature?

For example, technical ethical issues. The famous tram problem will reappear again. Suppose a self-driving vehicle faces a pedestrian who suddenly rushes out from the road ahead. It can perform evasive operations to protect the pedestrian, but it will sacrifice the safety of passengers and other vehicles on the road, and vice versa. Pedestrians, how should AI make judgments and choices?

There are still many problems faced by the development of autonomous driving, and we cannot list them all here. Because of this, companies committed to autonomous driving technology cannot act too hastily and promote some functional points that have not yet been implemented, which can easily cause public misunderstandings and even cause major safety accidents.

Automotive supplier Continental conducted a survey on self-driving in 2013. The results showed that 66% of Americans believed that “self-driving cars scare me” and 50% believed that “the technology Unable to operate reliably". By 2018, the data from the two surveys had increased to 77%. The reason may be that Tesla, Uber and other companies have experienced traffic accidents again and again during self-driving tests, which has affected public confidence in self-driving.

Baidu’s future, the future of autonomous driving

Robin Li predicted at the 2020 Baidu World Conference that autonomous driving will be fully commercialized in 5 years, urban congestion will be greatly alleviated, and purchase restrictions will no longer be needed. Traffic restrictions, and with the gradual popularization of driverless vehicles, the incidence of traffic accidents will be greatly reduced. The construction of intelligent transportation infrastructure based on vehicle-road collaboration will increase traffic efficiency by 15%-30%, thus contributing 2.4% to GDP. %-4.8% absolute growth.

Such business expectations are also supported by the market. RoboTaxi’s business prospects have been favored by many industry insiders and institutions.

McKinsey & Company predicts that China will be the world's largest autonomous driving market. By 2030, total sales of autonomous vehicles will reach US$230 billion, and orders for travel services based on autonomous driving will reach US$260 billion.

Market research firm iResearch predicts that as core technologies such as artificial intelligence, big data, cloud computing, and 5G mature, autonomous driving is entering a stage of rapid development. By 2022, the global autonomous driving penetration rate will reach more than 50%, and by 2030, the global autonomous driving penetration rate will increase to 70%.

No one doubts the future of autonomous driving. Waymo uses a valuation of 100 billion to tell all players that autonomous driving is a huge cake. However, the long journey of autonomous driving requires funds, time and technology, and because the threshold is higher, it is not like the blooming of new forces building cars, but a competition between the real masters in the industry.

Despite the long-tail effect in technology, the relevant policies and regulations are still unclear, and the year for large-scale commercialization is unclear, as the application scenario of L4 autonomous driving that attracts the most attention and is favored by capital, RoboTaxi has already Become a popular track.

Baidu, Waymo, Tesla, Cruise, who will lead this war? ?

This article comes from the author of Autohome Chejiahao and does not represent the views and positions of Autohome.