Mobileye, a part of Intel, is a leader in automation technology and the world's largest supplier of advanced driver assistance system (ADAS). Based on years of success in automotive automation and the experience that has evolved from ADAS to fully autonomous driving technology, Professor Shashua and his colleague, Shai Shalev-Shwartz, developed and published mathematical formulas that can bring certainty to questions about responsibility and criticism when accidents occur. This mathematical formula was also published at World Knowledge Forum in Seoul.
Their Responsibility Sensitive Safety model (RSS Model) provides specific and measurable parameters for the human concepts of responsibility and attention. It also defines a 'Safe State' in which autonomous vehicles can not be a cause for accidents, regardless of the movement of other vehicles. In order to verify the stability of autonomous vehicles, it is impossible to prove the high level of safety requirements required for autonomous vehicles. To solve this problem, it is necessary to introduce and apply a model-based proof method.
Amnon Shashua, CEO of Mobileye as well as vice president of Intel, pointed out the aspects of safety guarantee and economic expandability as fundamental issues in the difficulty of moving from a project to develop an autonomous vehicle to a mass production. In order to go to the stage of mass production, it is necessary to clarify the definition of 'stability' in society and to suggest measures for guaranteeing this stability. In terms of economic expandability, autonomous vehicles should be used anywhere in the world, and policies such as economic expansion and government incentives should be considered, which is necessary to make autonomous vehicles a single industry.
The safety problems of autonomous vehicles are conflicting with the two requirements. First, it is pointed out that autonomous vehicles, required to try 'negotiate' with their surroundings according to their circumstances and operate like humans, not conservative principles. For example, in situations of joining other roads or having heavy traffic, it is not wanted to be stuck up by autonomous vehicles, but to go through the situations at the same level as humans, while not to make an accident like humans.
In addition, autonomous vehicles can not have any accidents at all. Expectation could be lowered to the extent that accepts a little accident, but it is also necessary to identify the responsibility for the accident. If the automobile industry is responsible for the accident of an autonomous vehicle, the movement power to the autonomous vehicle may fall. And so far, the safety of automobiles has been dealt with at the hardware and system integrity levels, but autonomous vehicles need to be discussed at different levels. Autonomous vehicles require high level of decision making and sensing, and it is necessary to review the errors in sensing and decision making situations.
Verification of the safety of the vehicle is conducted on a statistical basis, and the maturity of the system is being verified with an empirical method that acknowledges reliability if there is no accident in the driving of millions of kilometers. However, this method is wrong for autonomous vehicles, and if it is continued to be used, autonomous vehicles may not be accepted in society and may disappear in the future. Also, the biggest contradiction in verifying autonomous vehicles by conventional means is expectation statistics about stability expected of existing cars and autonomous vehicles.
For example, in the United States, the number of deaths from traffic accidents is estimated to be around 35,000 a year. If autonomous vehicles are introduced at an annual death rate of 35,000, autonomous vehicles will not be accepted. The death of a person due to problems with vehicles and computers is socially unacceptable, and in order for receptiveness of the number of deaths from autonomous vehicles by the society, a very high level of achievement is required. Therefore, it was introduced that about 35 people, which is improved about a thousand times a year, and an achievement as the same level of aircraft could be accepted.
In order to prove these figures by conventional methods, if an autonomous vehicle travels about 30 km/h, it can be achieved by driving 30 billion km. Not only the distance is unrealistic, but also in terms of data collection and analysis, an autonomous vehicle that generates 5TB of data per hour will require 5 million PB of storage space. On top of that, when driving 20 hours per day for a year with 4 million vehicles that cost $100,000 per vehicle for driving distance of 30 billion km and considering costs for test drivers, it becomes unrealistic in terms of cost.
Likewise, while the existing data-based method for verifying the safety of autonomous vehicles has difficulty, the 'model'-based verification method has been chosen as an alternative to overcome this problem. It is also introduced that it will interpret and explain about situations. Moreover, it is based on the premise that it cannot guarantee a "complete safety", because the accident cannot be avoided in circumstances surrounded by uncontrollable factors, such as when the vehicle is surrounded by other vehicles. Also, it is because the purpose of the accident investigation is usually to cover responsibilities.
The Responsibility Sensitive Safety model provides specific and measurable parameters for the human concept of responsibility and attention. It also defines a 'Safe State' in which autonomous vehicles cannot be a cause for accidents, regardless of the movement of other vehicles. Also, during development of this model, the rules based on the common sense about the rules that could be matters of responsibility will be pre-defined. In this part, the detailed definition about matters of responsibility and the parts of model construction need to be discussed with regulation authorities. On the basis of this RSS model, the vehicle will control the behavior so that it moves only within the safety state, thereby enabling the safe driving without the accident responsibility of the autonomous vehicle.
This 'Safe state' allows vehicle to understand the speed, the road condition, the speed of the surrounding vehicles, and calculate the safe keeping distance of the vehicle. Even if the vehicle in front is suddenly stopped, the vehicle will maintain the distance regardless of reasons and enables driving without accidents. This 'safety distance' is set fairly conservatively, but it can be pulled more than this in reality. In the case of autonomous vehicles, it is possible to react quickly to changes in the surrounding environment, and the safety distance can avoid accidents when the car in front is suddenly stopped, even if it is 5.5m instead of a few hundred meters.
It is introduced that the RSS model could be applied in a case of extending autonomous driving scenario into a more complicated situation. It is also negotiable with authorities about rules, and it can establish methods for accurately identify matters of responsibility in a mathematical way. It is also capable of handling both-way traffic and signal waiting, and it will be able to respond to any situation that can be encountered on the road through a comprehensive related model. On the other hand, there is a need to discuss with the regulatory authorities the definition of safe state and the parameters that ensure that the autonomous vehicle does not become a cause for accidents in situations of driving like human with an aggressive inclination.
In conclusion, unlike an empirical model or a simulator, the RSS model was introduced as a model for ensuring safety. This section needs to be discussed with regulatory authorities, and the certification section of the vehicle also needs to be discussed. In addition, it is emphasized that the introduction of the RSS model will not only benefit specific companies but also help everyone in the autonomous vehicle-related industry. It will help them move to the mass production stage of an autonomous vehicle. The model, meanwhile, will be a publicly available model, and it is added that industry-wide and regulatory authorities should work together to help everyone in their ecosystem.
Copyright ⓒ Acrofan All Right Reserved