Researchers demonstrate learned bipedal maneuvers on a quadruped robot

Researchers report in an arXiv preprint that their reinforcement learning method controlled a physical quadruped robot as it rose onto two rounded rear feet, walked while balancing upright and returned to the ground. The method learns a linear representation of the robot's unknown nonlinear dynamics. It combines a one-step prediction with an estimate of future costs to improve the control policy, which then runs without online optimization. The researchers also report improved performance and sample efficiency against model-free and model-based baselines in locomotion simulations. Those comparisons come from simulations; the physical demonstration establishes the sequence of robot maneuvers.