Quadruped Robot Get Bionic Learning Method Based on Intelligent Memory Soft Actor-Critic

Quadruped robot gait control is a widely studied topic, with the advancements in artificial intelligence, reinforcement learning-based approaches provide a promising solution for bionic gait learning of quadruped robots. In this study, we propose an Intelligent Memory Soft Actor-Critic (IM-SAC) algorithm that uses the Soft Actor-Critic algorithm as the basic framework and incorporates Long Short-Term Memory network (LSTM) and Gated Recurrent Unit (GRU) to extract time-sequence-related sequence information of the quadruped robot motion. The IM-SAC algorithm aims to maximize the motion reward by controlling the degree of memory and forgetting of sample information, giving priority to learning samples with high reward values, and achieving faster cumulative rewards and optimization models. We design a reward function using the quadruped robot's own speed, swing angle, and other information to train the robot, and use the 12 motor angles as output values to control the movement of the quadruped robot. We conducted experiments on the Pybullet platform to test the algorithm's performance in gait learning tasks of the quadruped robot. The results show that our study provides a promising solution for gait control of quadruped robots by integrating reinforcement learning and intelligent memory mechanisms.