Skip to Main Content
In this research, an approach to optimize motions for a humanoids is presented. Rapidly-exploring Random Tree(RRT) were used to plan an initial suboptimal motion. A reinforcement learning was then implemented to optimize the trajectories with respect to energy consumption, similarity to a human's natural motion and, physical limits. Energy cost was estimated by joint torque from a dynamic model, and validated against actual measured torque values using system identification (SID). With a motion capture system, human motions were collected for a given set of tasks, producing a representative “natural” motion, another cost for optimization. Physical limits of each joint ensured spatial and temporal smoothness of generated trajectories. Finally, an experimental evaluation of the presented approach was demonstrated through simulation using MiniHubo model in OpenRAVE.