The thesis “Locomotion, step planning, and fall resistance in humanoid robots via reinforcement learning policies” studies how to equip humanoid robots with reliable walking and rapid recovery after a fall, without resorting to heuristics or preprogrammed trajectories. It addresses a central challenge in robotic autonomy: enabling real robots to act robustly in uncertain, contact-rich environments, while respecting the constraints of embedded computing. Purely model-based approaches show their limitations in terms of adaptation, while deep reinforcement learning (DRL) promises generalizable behaviors learned from data. The problem is therefore: how can we design DRL policies that are computationally light, transferable from simulation to real robots, and integrable into standard locomotion stacks? Methodologically, the thesis establishes the foundations of reinforcement learning applied to robotics, then proposes two main contributions trained in simulation with domain randomization and validated on small humanoids. FootstepNet is an efficient actor-critical step planner capable of producing continuous step placements while anticipating the number of steps needed to reach multiple local goals; it eliminates dependence on discrete step sets and heuristics, operates using embedded inference, and matches or exceeds the quality of ARA* planning, at a much lower computational cost, with validation in simulation and on a real robot at RoboCup 2023. FRASA, meanwhile, is a unified catching-up and lifting agent: a single policy transforms proprioceptive observations into motor commands that establish stabilizing contacts before standing up. By exploiting the Cross-Q algorithm and the robot's symmetry, FRASA reduces training to approximately 30 minutes and transfers zero-shot to the actual robot, surpassing a reference based on preprogrammed trajectories and handling a wide variety of initial postures. In conclusion, this work shows that lightweight, modular, and safe DRL policies can be made practical for onboard control of humanoids, significantly reducing downtime after disruption and paving the way for more general and robust autonomy in real-world conditions.