CPG BASED RL ALGORITHM LEARNS TO CONTROL OF A HUMANOID ROBOT LEG

Autonomous humanoid robots equipped with learning capabilities are able to learn tasks such as sitting down, standing up, balancing, walking and running. In this paper, central pattern generator (CPG) based reinforcement learning (RL) algorithm is applied to a robot leg with 3-links to balance it at upright by reducing dimensionality of the learning problem from 6 to 2. MapleSim is used for the leg modelling and this model is combined with the CPG based RL algorithm by utilizing Modelica and Maple software properties. Maple multi-body analysis template and Modelica custom component template allow symbolic inverse kinematics solution for the leg to be obtained. Thus, time and information lost in case of using a numerical solution are eliminated. The learning results show that the value function is maximized, temporal difference error is significantly reduced to zero and the leg is balanced at upright.