For most applications in the service industry, biped robots have advantages over other legged or wheeled robots to operate in human-centric environments. As they transition from the research lab to the real world, however, biped robots experience major uncertainties in unfamiliar environments. To reach their full potential in these domains, biped robots must be both robust to disturbances and energetically efficient. Underactuated dynamic biped robots exploit the natural dynamics of human locomotion to achieve highly efficient gaits. While practical for tasks requiring long autonomy, such robots face complex control challenges that stem from their unactuated dynamics and limited ability to reject large external disturbances in the unactuated degrees of freedom (DOFs). Some groundbreaking advances provide glimpses of natural agility and efficiency, but practical challenges still limit the ability of dynamic biped robots to achieve their envisioned potential.This dissertation uses the hybrid zero dynamics (HZD) control framework to develop TROPIC, a gait optimization package that unites the motion planning and controller synthesis problems. HZD-based control, however, is susceptible to velocity errors when the robot is pushed off the periodic orbit, so the resulting gaits are sensitive to perturbations in the unactuated DOFs. Robust bipedal gaits are generated by optimizing the dynamic coupling between the actuated and unactuated DOFs. The dynamic coupling quantifies the amount of control authority the robot has over the unactuated DOF, which is shown to be strongly correlated to robustness. The HZD formulation is also enhanced using heuristics derived from transverse linearization feedback control. The heuristics improve disturbance rejection performance of underactuated biped robots across a range of perturbations while maintaining simplicity of implementation for practical robots. The resulting controller is tested extensively in simulation and validated in experiment on the five-link biped robot ERNIE. Finally, robustness is improved using a novel gait switching method that enables dynamic biped robots to switch gaits autonomously following unexpected disturbances. While the main focus of this dissertation is on robustness improvement for efficient dynamic biped robots, this work also addresses another challenge that limits their practicality: scalability to high-dimensional systems.