Multi-Stage Deep Reinforcement Learning-Based Design and Optimization of Walking policy for Humanoid Robots

With the continuous advancement of research on humanoid robots, the optimization of motion control strategies has become a crucial research direction in this field. Walking, as the fundamental motion for humanoid robots to accomplish various tasks, its control policy directly affects the practical application of humanoid robots. Currently, the problems of low efficiency in policy training and high energy consumption in motion have gradually become prominent, which are key challenges to be solved urgently. To address this, a multi-stage deep reinforcement learning (DRL) framework is proposed for the generation and optimization of walking motion control policy. By sequentially solving the problems of motion policy generation and optimization, the results of simulation verify that the proposed pipeline can accelerate the policy training convergence and effectively improve motion energy efficiency under different walking speeds.