Collaborative Research: Human-Inspired Cyber-Physical Security in Human-Robot Interactions

This grant supports research on the safety and security of intelligent robots, advancing the safe deployment of robotic systems in manufacturing, healthcare, transportation, defense, and other public environments. Intelligent robots are increasingly expected to work near people and respond to physical contact, changing surroundings, and unexpected behavior. However, most existing human-robot interaction research assumes that people will cooperate with robots. As such, most robot security research represents an attacker as an abstract computer algorithm or an external disturbance. These approaches do not adequately explain the possibility that instead a person may observe a robot, identify a weakness, interfere with it, and change strategies in response to the robot’s behavior. This award supports fundamental research to understand and defend against adversarial physical interactions between humans and robots. The research will develop models of human decision making, methods for identifying realistic robot vulnerabilities, and control strategies that allow robots to resist and recover from disruptive actions. The results will contribute to the safer use of mobile robots of all kinds and may inform future testing methods and safety standards. The research involves several disciplines, including human factors, cognitive science, robotics, artificial intelligence, control theory, and cybersecurity. This multidisciplinary approach will support graduate and undergraduate education, mentored research experiences, and outreach activities for K–12 students. Current approaches to robot safety and resilient control do not account for a cognitively able human who selects and adapts physical attack strategies while interacting with a responsive robot. This research aims to fill this knowledge gap through three integrated activities that will develop: 1) a Cognitive Adversary Model to represent how humans perceive robot vulnerabilities, select disruptive actions, learn from previous outcomes, and update their strategies as the robot responds. The model will be developed and tested using immersive virtual-reality experiments that measure participants’ attack choices, gaze behavior, body movements, contact actions, physiological responses, and task performance; 2) a structured library of physical, perceptual, and environmental attack actions and combine the human cognitive model with AI-based search. Human participants will evaluate and refine AI-generated attacks, producing a set of effective attacks that remain physically and cognitively realistic; and 3) the team will use these attacks to develop robot controllers through offline adversarial training and adaptive co-training between robot defenders and simulated human-inspired attackers. The controllers will be evaluated in simulation, virtual reality, and a small-scale physical human-robot interaction study. The research will test the hypotheses that cognitive workload affects attack planning, that people update attack strategies based on observed robot responses, that human-constrained AI search discovers more realistic vulnerabilities than unconstrained search, and that controllers trained against human-inspired attacks are more resilient than controllers trained against generic disturbances. 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.