Fig. 3: Robustness of the system in the presence of perturbations and environmental changes.

a–e, Illustration of how the robot completes the task in perturbation test scenarios that it has not encountered during training and demonstrates the robustness of our proposed method. From left to right, we place 5 kg box-obstacles in front of the robot (a), push the robot from the front (b), alter the floor with an inclined (c) and slippery (d) surface and lesion the robot by removing the right foot (e). f, Sagittal and lateral CoM movement is shown under different perturbations demonstrating the amortized control. g, Hip pitch joint movement, which has the biggest effect on the motion during biped locomotion. The hip pitch joint motion is used to show how the policy adapts to the perturbation and rapidly re-executes a motion to counteract the perturbation.