Push recovery and active balancing for inexpensive humanoid robots using rl and drl

Push recovery of a humanoid robot is a challenging task because of many different levels of control and behaviour, from walking gait to dynamic balancing. This research focuses on the active balancing and push recovery problems that allow inexpensive humanoid robots to balance while standing and walking, and to compensate for external forces. In this research, we have proposed a push recovery mechanism that employs two machine learning techniques, Reinforcement Learning and Deep Reinforcement Learning, to learn recovery step trajectories during push recovery using a closed-loop feedback control. We have implemented a 3D model using the Robot Operating System and Gazebo. To reduce wear and tear on the real robot, we used this model for learning the recovery steps for different impact strengths and directions. We evaluated our approach in both in the real world and in simulation. All the real world experiments are performed by Polaris, a teen-sized humanoid robot.