Concurrent Training of a Control Policy and a State Estimator for Dynamic and Robust Legged Locomotion

In this paper, we propose a locomotion training framework where control policy and state estimator are trained concurrently. The consists of network which outputs the desired joint positions estimation estimates robot's states such as base linear velocity, foot height, contact probability. We exploit fast simulation environment to train networks transferred real robot. capable traversing diverse terrains hill, slippery plate, bumpy road. also demonstrate that learned can run at up 3.75 m/s on normal flat ground 3.54 plate with coefficient friction 0.22.