• Research

Our research addresses resilient and secure control of intelligent autonomous systems in uncertain and adversarial environments.

Research Projects

NIRVANA: Neural Implicit Rendering from Images at Varying Altitude with Nimble Algorithms

We aim to collect real data for the WRIVA project from ground level, aerial, and satellite imaging platforms of 8 200mx200m sites ranging from urban to rural, and develop and test a baseline structure from the motion pipeline from aerial/ground level and extend it to satellite imagery.

agents

An Observability Analysis Approach for Resilient Vision-Guided Unmanned Aerial Vehicles

Collaborative NSF Project with UNR: For multi-agent vision-guided dynamical systems, this research provides a framework which characterizes stealthy attacks based on the unobservable subspaces of both physical system dynamics and the neural network model used for perception. The proposed research advances state of the art by developing an online machine learning framework to learn control policies from image frames in real-time, by extending the adversarial neural network approaches for static images, and considering a simultaneous attack on the mission planning and control layer, perception layer, and visual sensors. Using stochastic optimization and simulation, the proposed attack detection methodology extends the existing model-based observer methods for linear time-invariant systems to deal with stealthy attacks against a networked system with autonomous agents represented by general time-varying and nonlinear models. Related Publications: [arXiv][2023 IEEE RA-L]

agents

Stealthy Attack Detection in Multi-Agent Control Systems

Stealthy cyber-physical attacks are challenging to detect and deleterious to the system stability. Stealthy attacks, however, can be detected by leveraging redundancies in system measurements or structure. Building on redundancy and reconfiguration, this project proposes a multi-level attack detection framework for multi-agent control systems. The level-based structure allows for the scalability of the approach and the privacy preservation of agents' state information. Related Publications: [IEEE CDC 2021][2022 ICUAS].

Safe Autonomy: Novel Autopilot Controllers for Improved Safety of Autonomous Systems

This project addresses autonomous emergency landing for aerial drones under system failures/attacks and other environmental uncertainties such as severe weather conditions using a Multi-Level Adaptation approach. This novel autopilot framework enables a drone under large uncertainties to perform safety maneuvers traditionally reserved for human pilots with sufficient experience. From a mission control perspective, the framework replaces the traditional top-down, one-way adaptation between mission planning, trajectory generation, tracking, and stabilizing controller, with a two-way adaptation between planning and control to improve the stability and robustness of the system. In the simulation, we used X-Plane 11 high-fidelity flight simulator to test and validate the solution. Related Publications: [AIAA SciTech 2023]

about image

(Left) Multi-Level Adaptation Autopilot Framework. To address the emergencies due to attacks/failures, adaptation in high-level missions, local goals, and low-level control tasks (multi-level adaptation) is required. (Right) Automatic emergency landing trajectories under engine fail implemented on a high-fidelity flight simulator.

Stealthy Zero-Dynamics Actuator Attack Against Sampled-Data Cyber-Physical Systems

This project develops a detection approach based on multirate sampling to detect stealthy zero dynamics attacks. We integrated the L1 adaptive control with the Simplex fault-tolerant software architecture to design a high-assurance controller for a fast recovery from the attack. The video shows the trajectory tracking control of a quadrotor subject to a zero-dynamics attack.