Runtime Verification at the Edge (NSF SHF 2118179)

The environments we live in, autonomous technologies we are developing, and even our bodies, are now instrumented: limited-resource nodes collect large amounts of data in real-time to better track and explain their system’s and environment’s behavior. A 2019 Cisco study found that there are 28.5 billion networked devices and connections in the world. Within this massive ecosystem, one class of future critical applications stands out: software applications that use networked nodes to provide detection of safety risks in the system or its physical environment. Example applications that require such monitoring include fleets of autonomous vehicles, health-monitoring wearables, search-and-rescue, and climate monitoring. These applications are already transforming lives, but suffer from a lack of timely, reliable and energy-efficient tools to monitor their correct operation. The focus of this project is to provide precisely such a monitoring infrastructure. This will require overcoming several difficulties: first, the monitoring code must be automatically generated, rather than hand-written, as this reduces the likelihood of errors. The monitor must be able to deal with analog/physical signals produced by the observed phenomena, such as wave heights or temperatures. It must also deal with drifting clocks on the different nodes, which do not read the same moment in time. It must also be resilient to node crashes and malicious attacks. Finally, it must be distributed over the nodes, rather than centralized, since this is less prone to catastrophic failures.

In this project, we will be radically extending the reach of runtime monitoring to new and economically important edge applications. This will be achieved by implementing three research thrusts: (1) the first theory for distributed monitoring of continuous-time, asynchronous signals. The algorithms perform distributed optimization on the edge nodes themselves, thus eliminating the need for a central monitor. Convergence proofs establish soundness. The algorithms incorporate partial knowledge of signal dynamics, where available, to accelerate convergence. (2) A theory and algorithms for incremental monitoring, where intermediate calculation results are still usable by the application should some nodes crash. The monitors will also accommodate nodes that intentionally falsify their data. (3) A rigorous validation of the algorithms on realistic autonomous vehicles, to establish their performance within a full software stack and in the presence of real-world noise and failure conditions.

F1/10 RACECAR: Community Platforms for for Safe, Secure and Coordinated Autonomy (NSF CNS/CCRI 1925652)

F1/10 is a shared, open-sourced infrastructure for the development and validation of new approaches to autonomous perception, planning, control and coordination. This community platform will facilitate autonomous system research, education and bench-marking through the creation of a new class of high-performance autonomous racing cars, that are 1/10th the size of a real (Formula 1) car and can reach a top speed of 50mph. The goal is to enable a wide range of experimental research and education in Safe, Secure, Coordinated and Efficient Autonomy. Much of today’s research on Autonomous vehicles (AVs) is limited to experimentation on expensive commercial vehicles that require large teams with diverse skills and power-hungry platforms. Testing the limits of safety and performance on such vehicles is costly and hazardous. It is also outside the reach of most academic departments and research groups. Furthermore, little research has been devoted to the development of safety benchmarks and infrastructure for certifying autonomous systems, guaranteeing the safety of data-driven decision making, power-efficient perception and control for efficient autonomy, active coordinated safety for large vehicle fleets, and cyber-physical security of autonomous vehicles. All of these are fundamental requirements on the road to achieving the social benefits of autonomous vehicles, and autonomous systems more generally. F1/10 provides rapid prototyping hardware, software and algorithmic platforms, with full documentation and community support for research and development of future autonomous systems.

Best Engineering Practices for Autonomous Systems (FAA ASSURE)

Industry desires to automate many types of Unmanned Aerial Systems (UAS) operations to the extent possible. This research will explore automation failures that provide insights relevant to the design of automation to support UAS operations. Researchers will classify the various types of automation failures that are identified in order to better assess automation risks and areas of design that require special attention to eliminate or mitigate these risks in the operation of UAS.

Operational constraints and limitations of UAS design will also be assessed as part of the overall risk management approach defining the conditions necessary for automation to safely perform during UAS operations. Generalized design guidance and best engineering practices will be developed. Specific design guidance and recommendations will be proposed for a few key areas of UAS automation design that require special attention, to include architecture, redundancy, runtime assurance, and self-monitoring. This specific guidance will be validated through simulations and experiments. Research findings will be shared with an industry standards body.

Fair Temporal Logic Control

When multiple robots share a resource, like UAVs, sharing the airspace, we intuitively do not want one of them to over-use the resource, forcing others to under-use it. This project looks at various notions of fairness and develops algorithms that maximize fairness while ensuring mission satisfaction. Without such fairness guarantees, small operators are unlikely to enter the market and economic benefits of new technologies, especially in Urban Air Mobility, are likely to be missed.