| Abstrakt: |
In the foreseeable future, the number of autonomous vehicles ("AVs") on the roads and highways will continue to rise, and these vehicles will become increasingly smarter, with some autonomous vehicles becoming truly "driverless." AVs will only be accepted on a wide scale when society trusts that they operate safely, without harming the passengers, other vehicles on the road, or bystanders. Because artificial intelligence ("AI") based software generally controls AVs, safety assessment and assurance of the AV software, although vital, is not straightforward. Existing laws, regulations, and international industry standards applicable to software included in other safety-critical systems do not adequately address the unique challenges that arise in the context of AV software because it is based on AI technology. Specifically, these regulations and standards generally emphasize specification-based functional testing and verification. However, the behavior and the safety requirements of AI-based software cannot be fully specified by the software designer. Therefore, the use of other testing schemes, such as scenario-based testing, is necessary. Scenario-based software testing is only effective when it is comprehensive, i.e., when a majority of the different situations or scenarios the software may encounter during the course of its operation in real life are contemplated beforehand and the software is tested using such scenarios. Driving presents a particular challenge in this context because the scenarios or situations a vehicle may encounter are virtually limitless. The quality and efficacy of scenario-based testing of AVs can, however, be greatly improved by sharing testing data among various entities involved with AVs, including makers, users, insurers, and regulators. To this end, I propose a new "Sharing fOr Safety" ("SOS") framework under which AV makers, developers of AV software, and other entities working with AVs can share the scenarios they individually contemplate or experience with each other. A sharing structure can make all AVs safer via access to a richer set of scenarios from which different AI-based AV software systems can learn and improve. Any concerns about losing a competitive advantage resulting from such sharing are counterbalanced by the lower risk of liability for accidents caused by AVs. Scenario-based testing, and sharing of scenarios according to the SOS framework, can be promulgated via laws, regulations, and industry standards, as well as agreements between entities, while simultaneously addressing user privacy concerns. The benefits of the SOS framework that I propose include lowering the risk of liability of the makers and users of AVs and, importantly, reducing the likelihood of harm to property and life. Both of these benefits stem from facilitating assessment and improvement of the safety of AI-based software controlling autonomous vehicles and, as a consequence, improving the safety of the vehicles themselves. [ABSTRACT FROM AUTHOR] |
