Collaborative Research: CPS-FR: PIAS: Performativity-aware Intelligent Autonomous Systems

Autonomous driving systems operate in environments that are not fixed. Their actions can change how people and other vehicles behave. For instance, when an autonomous vehicle slows near a crosswalk, it may encourage a pedestrian to cross; when it merges into traffic, it may cause nearby drivers to brake, accelerate, or change lanes. These responses then affect what the autonomous vehicle will experience next. Over time, such interactions can move the system away from the conditions in which it was originally trained, tested, and validated. We call this phenomenon "performativity." Performativity creates an important safety challenge for real-world autonomous driving. Current development practices often address new or unexpected driving situations by collecting more data, building new simulations, retraining models, and revalidating the system. While useful, this process can be slow, costly, and difficult to scale to the full range of complex traffic situations that autonomous vehicles may encounter. A more reliable approach is to design autonomous systems that can recognize how their own decisions reshape the driving environment and adapt accordingly. This project develops Performativity-Aware Intelligent Autonomous Systems (PIAS), a new framework that enables autonomous vehicles to anticipate, monitor, and safely respond to behavior changes caused by their own actions. By predicting how the environment will respond, planning actions safely within this evolving landscape, and learning stable models of behavior under changing policies, PIAS maintains reliability even as autonomy reshapes its surroundings. The resulting technology promises safer and more scalable autonomous driving, supports broader applications in intelligent systems, and advances education and workforce development in the rapidly growing field of autonomous technologies. This project advances knowledge in cyber-physical systems (CPS) by introducing a new paradigm for autonomy that explicitly models how control policies reshape future observations through policy-induced behavior. The core scientific contribution is the development of policy-induced distribution maps that enable autonomous systems to reason about how their actions influence the distributions they will subsequently encounter, addressing a critical and underexplored feedback loop in real-world autonomy. The proposed PIAS framework connects behavioral prediction with safe control so that autonomous vehicles can operate reliably in environments reshaped by their own actions. The research includes three integrated thrusts. First, it will develop behavioral models that learn how vehicle states, control actions, and environmental context influence future traffic behavior. Second, it will design control policies that use these predictions while explicitly accounting for uncertainty, especially in rare or unfamiliar driving situations. These policies will aim to provide probabilistic safety guarantees, meaning that the system can estimate and limit the chance of unsafe outcomes under specified uncertainty conditions. Third, the project will study how stable behavioral models remain when control policies change the environment. Meantime, methods will be developed to make these models more robust to such policy-induced shifts. Together, these efforts create a two-way framework in which prediction informs control, and control accounts for its effect on future behavior. The approach will be evaluated through policy-induced simulations and sub-scale vehicle experiments. The research results will contribute to safer, more reliable autonomy in complex, interactive traffic environments and provide a general methodology for behavior-aware cyber-physical systems beyond autonomous driving. 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.