We present a computationally efficient method for online planning of bipedal\nwalking trajectories with push recovery. In particular, the proposed\nmethodology fits control architectures where the Divergent-Component-of-Motion\n(DCM) is planned beforehand, and adds a step adapter to adjust the planned\ntrajectories and achieve push recovery. Assuming that the robot is in a single\nsupport state, the step adapter generates new positions and timings for the\nnext step. The step adapter is active in single support phases only, but the\nproposed torque-control architecture considers double support phases too. The\nkey idea for the design of the step adapter is to impose both initial and final\nDCM step values using an exponential interpolation of the time varying ZMP\ntrajectory.This allows us to cast the push recovery problem as a Quadratic\nProgramming (QP) one, and to solve it online with state-of-the-art optimisers.\nThe overall approach is validated with simulations of the torque-controlled 33\nkg humanoid robot iCub. Results show that the proposed strategy prevents the\nhumanoid robot from falling while walking at 0.28 m/s and pushed with external\nforces up to 150 Newton for 0.05 seconds.\n