Acquisition of energy-efficient bipedal walking using CPG-based reinforcement learning

Although there have been much research on robot walking, the energy efficiency of central pattern generator (CPG)-based walking has not received much attention. This study proposes a novel method for acquiring energy-efficient CPG-based bipedal walking for a robot with knees and feet. In this method, we introduce a torque-free period for swing leg control into the swing leg control cycle. During this period, no torque is applied to the hip joint controller, and therefore no energy is consumed. When and for how long the torque-free period is inserted into the swing leg control cycle is adaptively acquired by reinforcement learning. Simulation experiments demonstrate the feasibility of our method. The energy consumed in acquiring walking is reduced by 40% compared with simple CPG-based walking without the torque-free period in the practical learning speed. Walking stability is maintained with respect to external disturbances on a level floor. Although the method is more unstable on slopes with the torque-free period, the torque-free-period can be adaptively eliminated to achieve stable walking on the slopes.