Federated reinforcement learning enables agents to collaboratively train a global policy without sharing their individual data. However, high communication overhead and probabilistic risk constraints remain critical bottlenecks. These issues become more complex when the number of federated agents increases. We propose algorithms to solve such challenges.