domeet webmaster

Autonomous driving: the glorious picture of “achieving autonomous driving in 20 years”

According to international standards, according to the degree of intelligence, autonomous vehicles can be divided into 5 levels: L1-assist driving, L2-partial autonomous driving, L3-conditional autonomous driving, L4-Highly autonomous driving, L5-Fully autonomous driving (unmanned). Generally speaking, “autonomous driving” refers to high-level autonomous vehicles with L3 and above.

From a global perspective, L1 / L2 autonomous vehicles have achieved mass production. However, the L4 level of self-driving cars that are stepping on the edge of “autonomous driving” is still in the test or small-scale application stage of “in a limited area and a limited population”.

“Many companies have been vocalizing over the past few years: ‘we are really close to auto-driving’ or ‘We will achieve autonomous driving in the next 20 years ‘However, this is not the case.’ In John Dolan’s view, L5 “autonomous driving” in the true sense is still very far away.

The problem comes from two aspects: Underlying technology and real application scenarios.

Technical attack: the “last 5%” long tail problem

“The academic consensus is that we are already close to” autonomous driving “in the highway scene, but we are far away from achieving autonomous driving in citiesWe are pretty far from having cars in cities. “

In the discussion of autonomous driving, the current autonomous driving is reflected in the driving experience: people no longer need to pay attention to roadblocks. However, this simple requirement actually requires an integrated system to meet, and it is extremely difficult to use a complete computer system to deal with problems that may occur in all practical scenarios.

In the view of Professor John Dolan, from 2006 to 2007, the research on autonomous driving delineated by the academic community has formed a relatively complete framework. The theoretical framework of academic and industrial circles is:

• High-level: mission planning

• Middle level: behavior planning

• The bottom layer: motion control (motion planning)

Prof. John Dolan pointed out:

• Completing these different levels of requirements requires different technologies and algorithms, and requires the overall technical architecture to complete system functions under large-scale deployment and achieve human-like driving behavior. The development of the existing technology that has been gradually improved has indeed raised the technical baseline: for example, GPS has improved the accuracy of positioning and path planning; the development of deep learning, especially computer vision, has provided stronger capabilities for autonomous driving. Support. The ceiling of the material is being broken together by the manufacturing industry and the car companies. However, as an integrated system, autonomous driving also requires the development of the efficiency, computing power and computing speed of the overall system hardware in the industry.

• Autonomous driving as a daily life closely related to peopleActivities, its overall technical implementation requires not only the efficient operation of algorithms and software, but also the accurate operation of interactive systems. In the middle layer of the technical architecture, the decision-making system has changed from narrow vehicle behavior decision-making to requiring driving. “ Not only can reasonably decide the current vehicle behavior based on the information output by the perception layer, but also understand and predict the car’s driving External environment, predicting emergencies . ” Human behavior, as factor X in all models, is becoming the most difficult and difficult subject to overcome in autonomous driving.

In autonomous driving, the underlying architecture and most of the technical issues have been resolved, and the remaining 5% of unresolved issues have gradually become the decisive player in restricting the development of autonomous driving . The “last 5%” long tail problem is scattered in fragmented scenarios, special extreme situations and human behavior that can never be predicted, and it has plagued the academic community and industry at different levels of algorithms, sensors, computing platforms and regulations.

Real application scenario: 100% worthless

In order to better achieve the last 5% span, many companies choose to add real experiments.

However, according to John Dolan, the cognitive difference between laboratory experiments and real application scenarios is huge: The accuracy and credibility of laboratory data and models is in reality The environment may be worthless: The simulation of the laboratory environment can only make the experimental environment simulate the real environment as much as possible, but always meet the experimental environment settings of the real scene; Algorithm and system accuracy are critical: Scientists must continuously optimize the deployment to make the reliability of the algorithm constantly improve, and the real situation is often: even experiments that meet 100% accuracy are not in real life. It must work.

In addition to the differences between the laboratory and the real scene, there are also differences in the real scene. “Even if a successful test of autonomous driving in one city does not mean that we can replicate the successful experience in another city. Different societies, cities, driving environments and driving styles (Driving style) will affect the operation of autonomous driving. This is why most companies always insist on real experiments in different scenarios and cities.

“Only after the” final 5% special case “of imprecise simulation of the algorithm is resolved, can we judge that we have achieved L5 level autonomous driving.” John Dolan IntroductionRoad.

Reliability: a core issue in the field of autonomous driving

In the field of autonomous driving, Safety and reliability are always the core issues.

A year after the Uber self-driving car accident, a new report released by the NTSB shows that: The hardware of Uber’s self-driving test car has no problems, and the design is flawed. Software caused the fatal accident . Uber ’s driverless car accident warns a number of driverless companies: Security The problem needs to be resolved before the arrival of L5.

At present, autonomous driving mainly relies on the learning mechanism . Machine learning will be an important support among them to analyze the real situation, but the research of this part has just begun.

Current theoretical introductions also include:

• Formal vertification (specification, development and verification) with mathematical tools. However, the actual situation of automatic driving is often much higher than

• Ad hoc message inspection (adhoc message), which is calculated from the information taken in time and included in the prediction mechanism.

Why is the push for security and reliability so urgent?

Because the development of autonomous driving is attracting public attention and attracting pressure from the outside world.

• The government is the most important external force. The issue of high-enough reliability is being emphasized by governments everywhere, and government departments control the right to test real-world scenarios.

• The moral and psychological issues in society cannot be ignored . “ 1.24 million people die each year in traffic accidents worldwide, of which 91% of fatal traffic accidents occur in low- and middle-income countries. Autonomous driving may not change the number of casualties in traffic accidents, but people ’s psychology is very different for the same casualty results after the application of autonomous driving. “ John Dolan said,” People want machines to make fewer mistakes than humans, even though such expectations are wrong. “
  

While long-distance trucking has become the best stage for autonomous driving, the safety and reliability of autonomous driving is also linked to transportation costs and transportation effectiveness : idealization The state is that self-driving trucks can improve the effectiveness of long-distance transportation and safely deliver goods to their destinations, rather than causing large cargo losses due to accidents.

This involves another dimension that emphasizes the reliability and safety of autonomous driving: Economic effects . The impact of autonomous driving is two-way: when autonomous driving penetrates the streets, the driver ’s survival and employment space will be affected; on the other hand, the services of autonomous driving are still expensive, and not everyone can pay for the service. Stability also affects the willingness of passengers to pay-passengers will think: Is an autonomous driving service that is not stable or safe for you worth my purchase?

Competition: Leading Position and Lagging Advantage

In terms of autonomous driving, the importance of academia is self-evident: academia is not only technical research, but also high-quality human resources.

As the “Autonomous Driving Huangpu Military Academy”, CMU’s list of autonomous driving partners includes both Uber and Argo autonomous driving technology companies, as well as large automotive companies such as GM. Graduate startups from CMU dominate half of autonomous driving startups- Prof. John Dolan led the CMU driverless car team at the 2007 DARPA Urban Challenge (Carnegie Mellon’s Tartan Racing Team), and in this team alone, many autonomous unicorns or emerging companies have emerged: including Argo AI founder Bryan Salesky; Aurora founder Chris Urmson; Cruise founder Kyle Vogt Zoox founder Jess