The past few years has witnessed a significant progress in the field of legged locomotion. This is mainly due to the availability of high-performance torque-controlled platforms as well as development of algorithms that scale to high-dimensional, hybrid and under-actuated systems. In this talk, I will present my recent research efforts mainly on the algorithmic side, with an emphasis on developing efficient predictive controllers that can be complemented with supervised/reinforcement learning for real-time execution in the real world. I will also share my perspective on the open problems that we still need to solve to have functional humanoid robots in the real world