Controlling a Bipedal Robot with Pattern Generators Trained with Reinforcement Learning*
Title: Controlling a Bipedal Robot with Pattern Generators Trained with Reinforcement Learning*
Authors: Christos Kouppas (Loughborough University); Qinggang Meng (Loughborough University); Mark King (Loughborough University); Dennis Majoe (Motion Robotics LTD);
Year: 2019
Citation: Kouppas, C., Meng, Q., King, M., Majoe, D., (2019). Controlling a Bipedal Robot with Pattern Generators Trained with Reinforcement Learning*. UK-RAS19 Conference: “Embedded Intelligence: Enabling & Supporting RAS Technologies” Proceedings, 16-19. doi: 10.31256/UKRAS19.5
Abstract:
Herein, the use of reinforcement learning and pattern generators for balancing a bipedal robot, is described. SARAH (Silent Agile Robust Autonomous Host) is an underactuated robot designed by Motion Robotics LTD and aims to become an everyday bipedal robot that has fast, humanlike response. By utilizing V-Rep simulator, a simulated model of the robot was constructed and controlled with pattern generators. Then, those pattern generators were optimized by using reinforcement learning and a neutral advantage function agent. The training results are presented through graphs with respect to training steps, to show how the parameters converge to the optimum values.