Tremendous improvements have happened in the last decade in building autonomous vehicles, as analytics show that the increase of a variety of unmanned vehicles. These increases have been driven by shifts in several areas, including sensing and actuation, computing, control theory, design techniques, and modeling and simulation,” said Sandeep Neema, program manager at DARPA.
“In spite of these improvements, deployment and widespread adoption of such systems in safety-significant DoD applications remain challenging and uncertain.”
Last year the Defense Department issued a report on the popular state and future of freedom that focused on the need for self-sufficient systems to have a great degree of trust.
The report claimed that in order for autonomous machines to be believed they must operate safely and predictably, particularly in a military context. But workers should also be able to tell whether a system is working reliably, and, if not, the method should be created in such a way to allow proper action can be taken. This is the purpose of Assured Autonomy.
“Historically, support has been addressed through design processes following uncompromising safety standards in development, and confirmed compliance through system testing,” said Neema.
“However, these models have been produced primarily for human-in-the-loop systems, and don’t continue to learning-enabled systems with superior levels of autonomy. The support approaches today are meant on the premise that the systems, once extended, do not learn and evolve.”
One approach to an assurance of autonomous operations that has lately earned attention, especially in the context of self-driving vehicles, is based on the idea of “equal levels of safety,” i.e., the independent system must be at trivial as safe as a similar human-in-the-loop system that it follows. The approach links known rates of safety events of manned systems amount of accidents per thousands of miles operated and handling physical trials to define the similar incident rate for autonomous systems. Studies and analyses show, however, that ensuring the safety of autonomous systems in this practice alone is restrictive, requiring millions of physical trials, perhaps crossing decades. Simulation methods have been developed to overcome the required number of physical trials, but offer very little faith, especially with respect to low-probability, high-consequence events.
Take your time to comment on this article.