A Neurodynamic Approach to Stabilization of a 10 DOF Biped Mechanism Using Reinforcement Learning

In this paper, we propose a Reinforcement Learning (RL) based approach to stabilizing and control of a biped robot. Bipeds have complex requirements of stabilization and gait planning in multiple degrees of freedom. A reinforcement learning strategy is presented in this work for the stabilization of a single leg as well as a biped to create a learned behavior of a system. In this paper, each leg of the Humanoid biped robot is approximated as a double inverted pendulum, and its static stabilization is studied in the sagittal plane. The equations of motion are derived using Lagrange’s formulation method. An equivalent Humanoid robot single leg and biped model developed in Gazebo. Through Robotic Operating System (ROS), a reinforcement learning based control algorithm was developed for static stabilization, and the simulation was carried out on the Gazebo model. A total of 1458 states are used for training the reinforcement learning algorithm.