In this talk I will discuss my group's recent work on how to make multi-robot systems more robust and scalable. Many higher level decision making and coordination tasks in multi-robot systems can be abstracted as combinatorial optimization problems. While these algorithms work in theory, they often fail in practice because the abstraction ignores the uncertainty that's inherent in the real world. I will discuss our recent work on risk-aware combinatorial optimization that allows a user to trade-off risk and reward. Recently, learning has emerged as a practical tool for robot planning. However, these methods are hard to scale to large teams of robots, especially when they are heterogeneous. I'll present some of our recent work on scalable learning for multi-robot teams. Finally, I'll complete the loop by showing how learning can be combined with combinatorial optimization. I'll present some ongoing work on differentiable optimization that gives us the best of both worlds.