Biped Locomotion - Improvement and Adaptation

An approach addressing biped locomotion optimization is here introduced. Concepts from Central Pattern Generators (CPGs) and Dynamic Movement Primitives (DMPs) were combined to easily produce complex trajectories for the joints of a simulated DARwIn-OP. A Reinforcement Learning Algorithm, Policy Learning by Weighting Exploration with the Returns (PoWER), was implemented to improve the robot's locomotion through exploration and evaluation of the DMPs' weights. Maximization of the DARwIn-OP's frontal velocity while performing several tasks was addressed and results show velocities up to 0.25 m / s. The Stability and Harmony metrics were included in the evaluation and both charateristics were improved by the PoWER algorithm. The results are very promising and demonstrate the approach's flexibility at generating or adapting trajectories for locomotion.