In this work, we introduced a novel hybrid reinforcement learning scheme to balance a biped robot (NAO) on an oscillating platform, where the rotation of the platform is considered as the external disturbance to the robot. The platform had two degrees of freedom in rotation, pitch and roll. The state space comprised the position of center of pressure, and joint angles and joint velocities of two legs. The action space consisted of the joint angles of ankles, knees, and hips. By adding the inverse kinematics techniques, the dimension of action space was significantly reduced. Then, a model-based system estimator was employed during the offline training procedure to estimate the dynamics model of the system by using novel hierarchical Gaussian processes, and to provide initial control inputs, after which the reduced action space of each joint was obtained by minimizing the cost of reaching the desired stable state. Finally, a model-free optimizer based on DQN (λ) was introduced to fine tune the initial control inputs, where the optimal control inputs were obtained for each joint at any state. The proposed reinforcement learning not only successfully avoided the distribution mismatch problem, but also improved the sample efficiency. Simulation results showed that the proposed hybrid reinforcement learning mechanism enabled the NAO robot to balance on an oscillating platform with different frequencies and magnitudes. Both control performance and robustness were guaranteed during the experiments.
Model-Free Learning of Optimal Deterministic Resource Allocations in Wireless Systems via Action-Space Exploration
