Recent advances in artificial intelligence (AI), perception, and decision-making theory have facilitated the development of autonomous agents (e.g., ground, and aerial robots) capable of collaborating on complex tasks such as delivery, search-and-rescue, transportation, and manufacturing, in unknown environments. To handle environmental uncertainty effectively, decision-making in multi-agent systems often relies on AI techniques, particularly reinforcement learning (RL). RL translates perceptual feedback and shared information into individual control decisions. However, RL-driven control strategies have been shown to be vulnerable to adversarial conditions, including small perceptual noise, compromised shared information, agent failures, and self-interested agents. Despite recent empirical efforts to evaluate and enhance the robustness of such strategies, obtaining theoretical guarantees remains an open problem and a crucial challenge for ensuring the safe deployment of AI-enabled cyber-physical systems (CPS). This project aims to tackle this challenge. Specifically, the project?s novelties are the design of certification, verification, and robust training algorithms for multi-agent RL (MARL) control policies. If successful, this project will establish the scientific underpinnings for certifiable robust AI-enabled CPS making them suitable for safety-critical applications in potentially adversarial environments. The broader impact encompasses societal applications, facilitating the deployment of CPS in disaster relief, manufacturing, and surveillance scenarios. Additionally, the project will contribute towards educational initiatives spanning K-12 through graduate-level.<br/><br/>To achieve this research goal, three tightly coupled research directions are being pursued. The first objective focuses on training robust MARL control policies and theoretically certifying their robustness against adversarial high-dimensional sensory inputs. The second one focuses on assessing and enhancing the robustness of MARL control policies against adversarial communications and unexpected agent failures/removals. Extending these ideas to non-cooperative settings, the third research direction is explores the existence and computation of robust equilibria. Developing these fundamental analysis and synthesis capabilities for CPS necessitates designing new theoretical results drawing from formal methods, machine learning, and control theory. These research findings will be demonstrated on physical CPS testbeds and photorealistic simulators, using aerial and ground robot platforms, with particular emphasis on manufacturing and disaster relief applications.<br/><br/>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.