CPS: Medium: GOALI: Enabling Scalable Real-Time Certification for AI-Oriented Safety-Critical Systems

In avionics, an evolution is underway to endow aircraft with 'thinking' capabilities through the use of artificial-intelligence (AI) techniques. This evolution is being fueled by the availability of high-performance embedded hardware platforms, typically in the form of multicore machines augmented with accelerators that can speed up certain computations. Unfortunately, avionics software certification processes have not kept pace with this evolution. These processes are rooted in the twin concepts of time and space partitioning: different system components are prevented from interfering with each other as they execute (time) and as they access memory (space). On a single-processor machine, these concepts can be simply applied to decompose a system into smaller components that can be specified, implemented, and understood separately. On a multicore+accelerator platform, however, component isolation is much more difficult to achieve efficiently. This fact points to a looming dilemma: unless reasonable notions of component isolation can be provided in this context, certifying AI-based avionics systems will likely be impractical. This project will address this dilemma through multi-faceted research in the CPS Core Research Areas of Real-Time Systems, Safety, Autonomy, and CPS System Architecture. It will contribute to Real-Time Systems and Safety by producing new infrastructure and analysis tools for component-based avionics applications that must pass real-time certification. It will contribute to Safety and Autonomy by targeting the design of autonomous aircraft that must exhibit certifiably safe and dependable behavior. It will contribute to CPS System Architecture by designing new methods for decomposing complex AI-oriented avionics workloads into components that are isolated in space and time.

The intellectual merit of this project lies in producing a framework for supporting components on multicore+accelerator platforms in AI-based avionics use cases. This framework will balance the need to isolate components in time and space with the need for efficient execution. Component provisioning hinges on execution time bounds for individual programs. New timing-analysis methods will be produced for obtaining these bounds at different safety levels. Research will also be conducted on performance/timeliness/accuracy tradeoffs that arise when refactoring time-limited AI computations for perception, planning, and control into components. Experimental evaluations of the proposed framework will be conducted using an autonomous aircraft simulator, commercial drones, and facilities at Northrop Grumman Corp. More broadly, this project will contribute to the continuous push toward more semi-autonomous and autonomous functions in avionics. This push began 40 years ago with auto-pilot functions and is being fueled today by advances in AI software. This project will focus on a key aspect of certifying this software: validating real-time correctness. The results that are produced will be made available to the world at large through open-source software. This software will include operating-system extensions for supporting components in an isolated way and mechanisms for forming components and assessing their timing correctness. Additionally, a special emphasis will be placed on outreach efforts that target underrepresented groups, and on increasing female participation in computing at the undergraduate level.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.