Intel's Mobileye recently published a paper that pointed out that the mathematical formula can be used to determine the automatic driving vehicle in the collision accident responsibility attribution to ensure the safety of self-driving; this touched more than one automatic driving vehicle field observer's nerves , "What is the problem?" How can an industry match the player's own definition of product safety? "" Can security and responsibility be confused? "
For this reason, EE Times asked a number of scholar experts, ranging from robotics, embedded computer systems to autopilot vehicle safety, and human-robot interaction; we asked these experts to analyze Amnon Shashua, CEO of Mobileye / Intel senior vice president , And Shai Shalev-Shwartz, vice president of mobileye technology, to discuss whether they agree with the views and the problems they find, and provide them with advice on the industry.
Facts have proved that scholars are almost one side to give Mobileye research results positive view, they praised the company to adhere to in the end, in the field of self-driving the most difficult problems.
Phil Koopman, a professor at Carnegie Mellon University in the United States, said: "Overall, I think it would be nice to see a preliminary and rigorous approach to the safety of driving a vehicle automatically A car must have some way to judge what can be done and what can not be done, so I admire the author's work in this direction.
Missy Cummings, a professor at Duke, director of the Humans and Autonomy Lab, agrees with the idea that "I really appreciate Mobileye's start thinking about these issues in depth."
However, both Koopman and Cummings professors believe that Mobileye's approach is only "the first step" in the real world of resilience - especially when autopilot vehicles must coexist and interact with human-driven vehicles - Is a great leap; for what may be safe for self-driving, Mobileye's definition must be able to adapt to the harsh reality of the world.
Koopman value the value of Mobileye's specific recommendations for self-driving safety, he said: "No one can make the perfect advice for the first time, but it does not matter, we will try many different ways to express and formulate self-driving Security until we find a practical solution. "
The two authors of Mobileye's discussion of "security" in the paper is to explain that their strategy is "proving safe, and in this sense does not lead to traffic accidents that are attributable to autopilot vehicles"; but for this Duke University Cummings pointed out that "provable safe" (provable safe) is not a new concept, and cited on the Internet can find a lot of published academic papers (refer to the link).
She said that the most difficult problem of security has not changed: "Computer scientists from a mathematical point of view to consider what can prove security, does not mean that with the test engineers will agree that it is safe."
Must be questioned
Both Koopman and Cummings cautioned by the assumptions made by Mobileye, saying that it can not be taken for granted and needs to be questioned; as Koopman points out, "there are some assumptions that if you really happen in the real world, I will be very surprised."
Cummings's example is a software error; the following is a description of the security issue by Mobileye's author:
... We are now discussing errors that lead to unsafe behavior; as mentioned earlier, our strategy is provable to be safe, and in this sense it does not lead to traffic accidents that are attributable to autopilot vehicles. (Such as all the sensors are damaged, or puncture on high-speed power), software failure (some models have serious errors), or sensing errors occur. Our ultimate goal Is to minimize the possibility of such events - the probability of 10 9 times per hour.
Cummings raised questions about the potential problems that Mobileye claims to be a potential problem caused by software errors, and she cited a report that explores the events of the history of vehicles that are frequently recalled by security concerns, often caused by software problems.
Koopman's doubts are the problems of the lidar and the radar: "It is hard to believe that the independence of Raytheon and radar can be resolved and the assumptions that are discussed." He pointed out that "someone must prove that they are right Is not just a hypothesis, and it is almost certain that some assumptions are wrong, and the authors do not even realize that they have made a mistake. "
It's like Koopman worries: "This is security - the accident is the hardest part, so you need to plan for those accidents, and be careful to notice them when they happen;" but he also said "I'm still glad to see those authors start making assumptions, and they know they are assuming that because of this, we have a starting point to test those assumptions."
Define security
Koopman is not very worried about Mobileye's definition of security: "The paper claims that if they have a harsh definition of attribution of responsibility, you can build a system that will never be blamed "He said:" From a system point of view, this can make it safe, and if every car on the road has such a strategy and will follow, they think everything will be quite safe. "
While admitting the worries about the actual implementation of this concept, Koopman once again stressed: "Knowing why it might not work is a very important part; and seeing a really concrete suggestion that we can think and learn from it Good things. "
What is the problem with Koopman that there are potential problems with the approach proposed by Mobileye? His doubts are that autopilot vehicles "may learn how to 'cheat' the system '; in the real world, human driving tends to find loopholes in road rules and then To some extent the use of these loopholes, then why a "artificial intelligence" of the self-driving can not play the same trick?
(To be continued)
Compiler: Judith Cheng