Current applications for humanoid robotics require autonomy in an environment specifically adapted to humans, and safe coexistence with people. Whole-body control is promising in this sense, having shown to successfully achieve locomotion and manipulation tasks. However, robustness remains an issue:​ whole-body controllers can still hardly cope with unexpected disturbances, with changes in working conditions, or with performing a variety of tasks, without human intervention. In this thesis, we explore how whole-body control approaches can be designed to address these issues. Based on whole-body control, contributions have been developed along three main axes:​ joint limit avoidance, automatic parameter tuning, and generalizing whole-body motions achieved by a controller. We first establish a whole-body torque-controller for the iCub, based on the stack-of-tasks approach and proposed feedback control laws in SE(3). From there, we develop a novel, theoretically guaranteed joint limit avoidance technique for torque-control, through a parametrization of the feasible joint space. This technique allows the robot to remain compliant, while resisting external perturbations that push joints closer to their limits, as demonstrated with experiments in simulation and with the real robot. Then, we focus on the issue of automatically tuning parameters of the controller, in order to improve its behavior across different situations. We show that our approach for learning task priorities, combining domain randomization and carefully selected fitness functions, allows the successful transfer of results between platforms subjected to different working conditions. Following these results, we then propose a controller which allows for generic, complex whole-body motions through real-time teleoperation. This approach is notably verified on the robot to follow generic movements of the teleoperator while in double support, as well as to follow the teleoperator’s upper-body movements while walking with footsteps adapted from the teleoperator’s footsteps. The approaches proposed in this thesis therefore improve the capability of whole-body controllers to cope with external disturbances, different working conditions and generic whole-body motions.