Though such situations may be rare, they cannot be ignored — simple arithmetic dictates that in order for failures to occur at least as infrequently as they do with human drivers, a system must handle many such rare cases without failure. For systems that rely on machine learning, the need to get rare cases right has implications for system design and testing.
Machine learning approaches can be more confident that a case will be handled correctly if a similar case is in the training set. The challenge is how to develop a data set that includes enough of the rare cases that contribute to the risk of an accident.
Commercial aviation has mechanisms for sharing incident and safety data across the industry, but reporting may not be second nature to recently credentialed UAS operators who are new to the safety and accountability culture of the traditional aviation industry.
No comparable system currently exists for the automotive industry — only fatal accidents are reported, and the collection and reporting of other traffic safety information is done, if at all, in a disparate manner at the state or local level. The lack of consistently reported incident or near-miss data increases the number of miles or hours of operation necessary to establish system safety, presenting an obstacle to certain AI approaches that require extensive testing for validation.
To facilitate safe testing, the FAA has designated six UAS Test Sites across the country and provided blanket authorization for UAS operations within these sites. Activities at the sites include a project to extend NASA’s multi-year research on UAS traffic management (UTM) to identify operational requirements for large-scale beyond visual line-of-sight UAS operations in low-altitude airspace.
Similarly, ground vehicle test beds such as the Connected Vehicle Pilots and the deployment of automated vehicles in Columbus, Ohio, winner of the Department of Transportation’s $40 million Smart City Challenge in 2016, will provide rich baseline and interaction data for AI researchers.
Adapting Current Regulations While the regulatory approaches for the Nation’s airspace and highways differ, the approaches to integrating autonomous vehicles and aircraft share a common goal: both the FAA and NHTSA are working to establish nimble and flexible frameworks that ensure safety while encouraging innovation.
With respect to airspace regulation, a significant step toward enabling the safe integration of UAS into the airspace was the FAA’s promulgation of the Part 107, or “Small UAS,” final rule, which took effect on August 29, 2016. For the first time, the rule authorizes widespread non-recreational flights of UAS under 55 pounds. The rule limits flights to daytime, at an altitude of 400 feet or less, with the vehicle controlled by a licensed operator and within the operator’s direct line of sight. Flights over people are not allowed.
Subsequent rules are planned, to relax these restrictions as experience and data show how to do so safely. In particular, DOT is currently developing a Notice of Proposed Rulemaking proposing a regime for certain types of “micro UAS” to conduct operations over people, with a rule on expanded operations expected to follow.
The FAA has not yet publicly announced a clear path to a regulation allowing fully autonomous flight. Though safe integration of autonomous aircraft into the airspace will be a complex process, the FAA is preparing for a not-so-distant technological future in which autonomous and piloted aircraft fly together in a seamlessly integrated airspace system.
New approaches to airspace management may also include AI-based enhancement of the air traffic control system. Projected future air traffic densities and diversity of operations are unlikely to be feasible within the current airspace management architecture, due to current limits on air/ground integration, and reliance on human-to-human communication in air and ground practices.
The cost of U.S. air transportation delays in 2007, the latest year for which there is reliable public data, was estimated to be $31.2 billion — a number that has presumably grown as user volume has increased since that year.
Though some flight delays are unavoidable due to weather and other constraints, adopting new aviation technologies, enabling policies, and infrastructure upgrades could significantly increase efficiency of operation in the U.S. airspace.
Such solutions include AI and machine learning-based architectures that have the potential to better accommodate a wider range of airspace users, including piloted and unpiloted aircraft, and to use airspace more efficiently without undermining safety.
Development and deployment of such technologies would help ensure global competitiveness for airspace users and service providers, while increasing safety and reducing cost. With respect to surface transportation, the most significant step currently underway to establish a common framework is the Federal Automated Vehicles Policy that the Administration released on September 20, 2016.
The policy had several parts: guidance for manufacturers, developers, and other organizations outlining a 15 point “Safety Assessment” for the safe design, development, testing, and deployment of highly automated vehicles; a model state policy, which clearly distinguishes Federal and State responsibilities and recommends policy areas for states to consider, with a goal of generating a consistent national framework for the testing and operation of automated vehicles, while leaving room for experimentation by states; an analysis of current regulatory tools that NHTSA can use to aid the safe development of automated vehicles, such as
- interpreting current rules to allow for appropriate flexibility in design,
- providing limited exemptions to allow for testing of nontraditional vehicle designs,
- and ensuring that unsafe automated vehicles are removed from the road; and a discussion of new tools and authorities that the agency could consider seeking in the future to aid the safe and efficient deployment of new lifesaving technologies and ensure that technologies deployed on the road are safe.
DOT intends for the guidance and the model state policy to be routinely updated as new data are learned and research completed.