Control of Coal Mine Quadruped Robot Based on Reinforcement Learning

FU Xin

doi:10.16545/j.cnki.cmet.2025.03.005

Abstract：In recent years,with the widespread application of coal mine robot inspection technology,wheeled,track and crawler inspection robots have become common configurations.However,compared to these traditional configurations,quadruped robots are capable of performing inspection tasks underground in coal mines due to their strong terrain adaptability,especially in dealing with various complex scenarios such as steps,tracks,and rugged gravel surfaces.Due to the complex terrain and slippery ground conditions of underground operations,the motion control and state estimation of quadruped robots face severe challenges.For this purpose,an end-to-end neural network controller was designed,which integrates a state estimation module,an electric motor adaptation module,and a control module to achieve accurate estimation of robot speed,pose,electric motor drive parameters,and real-time dynamic characteristics.In terms of implementation methods,a dynamic simulation environment was constructed based on a physics engine,and the neural network controller was trained through deep reinforcement learning.To improve the generalization ability of the controller,random domain techniques were used during the training process to randomize key parameters such as friction coefficient,motor dynamics,and robot body mass.In addition,asymmetric actor commentator reinforcement learning algorithms,privileged observation data,and recurrent neural network structures were introduced to further enhance the training effectiveness.The experimental results showed that the designed neural network controller exhibited excellent robustness in complex underground environments,significantly improving the motion control performance and state estimation accuracy of quadruped robots.