Cooperative machine learning is reshaping how multi-agent systems handle massive, distributed, and privacy-sensitive data streams. This chapter condenses my doctoral work on two fronts: learning, which builds shared models, and inference, which uses those models in real time. For learning, we propose graph-aware message-passing neural networks that surpass classical belief propagation in data association and joint positioning, fully decentralized federated and split schemes that protect privacy and save energy in medical and internet of things (IoT) settings, and a reinforcement framework that keeps localization accurate in highly dynamic vehicle swarms. For inference, we present a single-pass detector for instant non-line-of-sight recognition, a latent-feature fusion method that switches smoothly between standalone and cooperative static positioning, and a sampling-free Bayesian kernel that attaches trustworthy aleatoric and epistemic uncertainty to mobile tracking. Together, these contributions form a coherent toolbox that turns heterogeneous 5G/6G, robotic, and healthcare networks into reliable, low-latency cyber-physical systems.