Ubiquitous sensing enables large-scale multi-source data of cyber-physical systems (CPS) collected in real-time and poses both challenges and opportunities for a paradigm-shift to AI-enabled CPS. Existing networked CPS decision-making frameworks lack understanding of the tridirectional relationship among communication, learning and control, let alone of how to define and quantify formally the benefits of information sharing. In this talk, we first present our research contributions on learning and control based on information sharing for CPS. We design a novel uncertainty quantification method for collaborative perception of connected autonomous vehicles (CAVs), and show the accuracy improvement and uncertainty reduction performance of our method. To utilize the information shared among agents, we then a safe and scalable deep multi-agent reinforcement learning (MARL) algorithms with truncated Q-function and safety guarantee based on continuous state space controller. To motivate agents to coordinate, we design a stable and efficient Shapley value-based reward reallocation for MARL. Considering system state uncertainties or even adversarial perturbations, we analyze what is the solution of MARL under state uncertainties and design robust algorithm to learn robust policies. We then briefly present our research contributions on data-driven robust optimization for efficient and climate friendly mobile CPS. Finally, we will highlight our research results on CPS security.