Mitigating Faults and Cyber Threats in Autonomous Systems Through Verifiably Safe Control

Safety is a critical property of autonomous systems such as driverless cars, unmanned aerial vehicles (UAVs), surgical robots, and energy systems. Naturally occurring faults and deliberate attacks may lead to safety violations, for example, Global Positioning System (GPS) denial attacks that cause vehicles to crash into pedestrians, or disabled actuators that leave a robotic arm stuck in an unsafe position.

This award supports research that enables the development of a framework for safe control in the presence of sensor and actuator faults and attacks, thereby promoting the progress of science, advancing prosperity and welfare, and securing the national defense. This approach combines techniques from control theory, machine learning, and system security to enable provable guarantees across a wide range of autonomous systems and fault/attack scenarios.

The developed techniques will be especially valuable for learning-enabled systems, which are known to experience severe performance and safety degradations when they encounter situations (such as deliberate attacks) that did not occur in their training data. This timely effort will enhance safety and trust, and hence pave the way for widespread deployment, of autonomous and learning-enabled systems across a variety of application domains.

Beyond technical advancements, this project emphasizes a variety of education and outreach activities including course modules on fault tolerant control and data-driven method, and undergraduate capstone and Research Experiences for Undergraduates (REU) projects on cyber-resilience of autonomous vehicles, targeting towards underrepresented groups.

This research will be grounded on nonlinear dynamical systems experiencing (i) proprioceptive sensor attacks, which target sensors that estimate the system's position, velocity, and other internal states, (ii) exteroceptive sensor attacks, which target sensors that gather information on the surrounding environment, and (iii) actuator faults. A unifying framework is planned to address (i)-(iii).

In this framework, the system maintains a collection of safety filters, each of which constrains the control input at each time step in order to ensure safety under a particular failure or attack. Decision modules will also be developed for determining which safety filters are critical at each time step and which can be relaxed. Furthermore, exact conditions will be developed for formally verifying safety under multiple faults and attacks.

The formal approach will build on foundations from algebraic geometry to formulate safety verification as a convex optimization problem. This framework will be compatible with physics-based and data-driven (neural network) system models under a variety of faults and attacks. The techniques will be validated through simulation and hardware evaluation on the PI's F1tenth racing testbed.

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.