In addition to Weibo, there is also WeChat
Please pay attention
WeChat public account
Shulou
2025-04-08 Update From: SLTechnology News&Howtos shulou NAV: SLTechnology News&Howtos > Internet Technology >
Share
Shulou(Shulou.com)06/02 Report--
2019-08-28 20:39:38
Relevant laws and regulations are no longer the main bottleneck.
The development of self-driving car technology also makes governments around the world step up the formulation of laws and regulations related to it.
At the federal level, the U.S. Department of Transportation (DOT), in place of the U.S. Highway Traffic Safety Administration (NTHSA), updated and issued a manual on self-driving vehicles at the end of 2018, extending the concept of self-driving to all ground road vehicles, clarifying the regulatory responsibilities of federal and state governments, and focusing on eliminating unnecessary policy constraints that may affect the development of self-driving technology. At the state government level, 31 states have issued self-driving laws or local administrative orders by the end of March 2019, but there is no legal basis to support self-driving vehicles on the road in 19 other states, such as New Jersey and Rhode Island. Japan issued the Road Test Guide for self-driving cars in 2016, allowing road testing of self-driving cars, while South Korea amended the Motor vehicle Management Law in 2016 to allow self-driving car testing on the road. The Ministry of Industry and Information Technology, the Ministry of Public Security and the Ministry of Communications jointly issued the Intelligent Network connection vehicle Road Test Management Standard (for trial implementation), standardizing the test theme, vehicles, license application and so on. At present, Beijing, Shanghai, Guangzhou, Shenzhen and Chongqing have successively issued autopilot road test laws and regulations and issued special road test licenses.
Algorithms and data: self-driving companies begin to embrace open source datasets
At present, the key difficulty of L4 autopilot is that the existing algorithms can not accurately deal with infinitely possible long-tail scenarios in complex environments. Once the difficulties of the long-tail scenario algorithm are overcome, the safety concerns of autopilot will be greatly alleviated, and we believe that autopilot is expected to accelerate the process of commercialization. Commercialization will bring large-scale mass production of self-driving parts and vehicles, so as to reduce the cost of sensors and promote the rapid maturity of the industrial chain, and the current difficulties of perception layer and execution layer will be solved naturally.
From the perspective of perception level, decision-making level and executive level, we think that the difficulties of L4 autopilot landing include:
Perception layer: at present, the sensor technology of the perception layer has basically met the requirements of autopilot. However, sensors such as lidar are still too expensive for passenger cars. For commercial vehicles, the cost of lidar is not a major obstacle because of its operational nature. Understanding and decision-making level: the scene from closed to open, the larger and more complex the open environment, the challenge to the self-driving decision-making level will increase exponentially. The decision-making algorithm not only needs to accurately deal with the data of multi-sensor fusion, but also is responsible for vehicle path planning, behavior planning and trajectory planning. The current autopilot algorithm is difficult to perfectly deal with scenarios that have never been seen before, and there are numerous possibilities in open scenarios, so it is difficult for an enterprise to fully consider it in its own algorithm in just a few years. In fact, despite years of training, Waymo had problems turning at intersections where there was no left turn in 2018. Executive layer: the current self-driving test cars are all converted from existing vehicles, and their stability and reliability are still far away from the requirements of mass production cars. If the self-driving model is to be launched, the electronic and electrical architecture, throttle, steering, braking and other execution systems of the vehicle need to be redesigned, and the verification test will also take a certain period after the completion of the model design.
For self-driving enterprises, data is an important asset, and high-quality tagged data is very important for autopilot development. In the past, the vast majority of self-driving companies kept their data sets strictly secret, but now they are accelerating the shift to openness. In March 2018, Baidu Apollo first released the autopilot data set ApolloScape. In June 2019, at the top meeting of computer vision CVPR 2019, Waymo and Argo AI also released public autopilot data sets, in which Waymo Open Dataset labeled data of up to 600,000 frames and rich sensor configuration, while Argo AI's Argoverse data set was the first public data set containing high-definition map data. Then, in July 2019, Lyft released an open source autopilot dataset. We believe that, considering the current difficulties encountered by autopilot in long-tail scenarios, it is difficult for individual enterprises to independently establish and maintain perfect data sets, so the opening of autopilot data is a long-term trend. data openness will help the self-driving industry to break through the algorithm bottleneck.
Zero defect in the system is the core challenge to realize autopilot, and L3 may be the short-term optimal dynamic balance.
Road safety is one of the key elements in the process of realizing autopilot. We believe that the goal of zero defect pursued by a complex system such as autopilot is becoming its own constraint. High-level autopilot means the release of the driver's attention and time, and after excluding human factors, it also means deepening the dependence on the safety of equipment.
Raise the safety standard of the autopilot system to a higher position.
From the perspective of hardware capability, autopilot improves the complexity of the system, such as visual cameras, millimeter wave radar, lidar, high computing chips, complex algorithms and other fields, the mainframe factory also lacks mature experience. At the same time, the traditional body systems, such as airbag system, braking system, chassis control system, powertrain control system and related power electronic components, have been greatly upgraded because they have been brought into the category of generalized ADAS actuator. The increase in system complexity increases the risk of random failure of a single hardware or system-level failure. From the perspective of project development, it is necessary to ensure that the risks related to autopilot systems are predictable, quantifiable and traceable in the whole life cycle (management, research and development, production, operation, service).
ISO26262 identifies critical hardware and adds redundancy. According to the definition method of ISO26262, the risk degree of the autopilot system can be as follows:
When the danger occurs, the severity of the injury (severity), the probability of occurrence of the dangerous working condition (exposure rate), and the controllability of the hazard
Three indicators are considered, and it is divided into four levels of functional safety level ASIL A/B/C/D. ASIL-An is the lowest security level, ASIL-D is the highest security level, and QM means that the system requirements have nothing to do with security functions. The higher the ASIL level, the higher the system security requirements, but at the same time, it means that the higher the hardware diagnosis coverage, the more stringent the development process, resulting in higher development costs and longer development cycle.
The failure of most advanced autopilot functions (such as autonomous steering and emergency braking) may lead to fatal injuries to drivers, so it is more likely to be classified as the safety level of ASIL-D by the mainframe factory, which requires necessary redundancy for many key subsystems.
The weak ability of environment recognition is not the defect of hardware physics, but the deficiency of algorithm.
At present, it is still difficult to make an accurate judgment of the external environment at the level of autopilot technology. Subject to the diversity of the weather environment, the complexity of the road environment, the high-speed movement characteristics of the vehicle itself, and the sensors are affected by factors such as viewing angle, illumination, fouling, occlusion and so on, there is no ideal environmental sensing technology that can properly deal with all working conditions. it needs to be coupled with multi-sensing devices and algorithms to make up for the shortcomings of various technologies.
Therefore, the recognition ability of the environment is difficult to tend to zero defects, not because of the physical defects of the parts, but more because of the lack of the ability of the algorithm to deal with the complex environment.
Zero defect management is not free, which brings about the transfer of industrial profit pool in the process.
There is no doubt that product defects will increase the cost of the enterprise. The cost will include failure analysis, rework, re-inspection, quality assurance cost, recall cost, turnover extension and so on, so when the quality of the automobile system is poor and the product defect rate is high, the overall quality cost is very high. Takata airbag product defects caused many well-known vehicle companies around the world (Toyota, Nissan, Honda, Acura, Mazda, Ford, etc.) to recall more than 10 million vehicles, and directly led to Takata's bankruptcy in 2017; Toyota also brought about a recall fee of about $2 billion because of accelerator pedal failure in 2010.
However, zero defect management is not free, it needs to be integrated into the autopilot system research and development process, process, and even need to reshape the corporate culture.
From the system level, autopilot faces the optimization of system architecture and brings about the transfer of profit pool: from distributed system architecture to centralized system architecture, from many control units to a small number of centralized processors. By reducing the number of hardware, this is not only a change in hardware capabilities, but may also be accompanied by a shift in the value chain. In the past, the control unit of each subsystem of the distributed architecture could undoubtedly provide additional value and improve the profits of the enterprise, while the centralized architecture downgraded the parts supplier to a pure hardware supplier and concentrated the profit pool to the whole vehicle enterprise or the first-level system vendor. From the perspective of the industrial chain, it increases the cost of communication. When the autopilot system is designed to a large number of parts / subsystem suppliers, the zero defects of the system will eventually be decomposed into the defect rates of all subsystems and components: from various sensors to processing chips, from operating systems to actuators, what is needed is not only the contribution of the whole vehicle enterprise, but also the contribution of the whole industry chain to higher system security. From the point of view of functional safety, the development cost and material cost increase. Due to the need to establish a standardized development process to ensure that all requirements can be taken into account and traceability can be taken into account in the product development process, and that all designs can be detected and verified. Zero defect management process is also a process to increase development costs. The hardware redundancy required for functional safety directly increases the material cost.
L3 autopilot schedule is a dynamic balance between cost and defect rate.
For self-driving, blindly pursuing complete "zero" defects, in the short term, due to the lack of accuracy in areas such as environmental identification, if a large number of hardware is piled up, it will bring extremely high cost pressure. We believe that L3-level autopilot is the result of a dynamic balance between cost and defect rate, and handing over some of the results of system defects to the driver may be a necessary tradeoff before the real realization of "unmanned" driving.
Https://www.toutiao.com/a6730196480230949383/
Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.
Views: 0
*The comments in the above article only represent the author's personal views and do not represent the views and positions of this website. If you have more insights, please feel free to contribute and share.
Continue with the installation of the previous hadoop.First, install zookooper1. Decompress zookoope
"Every 5-10 years, there's a rare product, a really special, very unusual product that's the most un
© 2024 shulou.com SLNews company. All rights reserved.