Reinforcement Learning based Control of a Quadruped Robot

Quadruped robots are four legged robots having potential to roam almost in all the earth surfaces in different terrains. A quadruped robot which is passively stable with 12 DOF is considered in this paper. The 12 DOF quadruped model is developed in Fusion 360 and later imported to ROS. The mathematical modelling of the quadruped system is complex and conventional control requires the complete knowledge of the system. Hence, Reinforcement Learning based algorithm can be used for the control of a quadruped system. The steps involved in making the robot walk include stabilizing the robot at neutral position, planning footsteps, generating body trajectory, planning swing leg trajectory and solution of inverse kinematic model to find the joint angles. A deep deterministic policy gradient (DDPG) agent is used to train the system to walk in straight line. The DDPG agent uses a critic value function representation to approximate the long-term reward given observations and actions. This quadruped is capable of walking amidst disturbances unlike a quadruped with conventional control.