Sim-to-real: Quadruped Robot Control with Deep Reinforcement Learning and Parallel Training

In recent years, deep reinforcement learning methods provide a new realization idea for the motion control of the quadruped robot, which makes the neural network controller have the characteristics of strong adaptability and high stability. In this paper, we propose an end-to-end neural network fram...

Full description

Saved in:
Bibliographic Details
Published in2022 IEEE International Conference on Robotics and Biomimetics (ROBIO) pp. 489 - 494
Main Authors Jiang, Han, Chen, Teng, Cao, Jingxuan, Bi, Jian, Lu, Guanglin, Zhang, Guoteng, Rong, Xuewen, Li, Yibin
Format Conference Proceeding
LanguageEnglish
Published IEEE 05.12.2022
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:In recent years, deep reinforcement learning methods provide a new realization idea for the motion control of the quadruped robot, which makes the neural network controller have the characteristics of strong adaptability and high stability. In this paper, we propose an end-to-end neural network framework that, contains an estimator network for estimating the robot's ontology states in addition to the critic and actor networks. These states serve as an important observation data for the critic and actor networks. In addition, in order to solve the disadvantage of high cost of neural network training time, we exploit the mechanism of parallel training and deploy the entire training process in GPU, which improves the speed of network convergence. Finally, we reconstruct the forward network on the CPU through the Eigen library, and transfer the model to the real robot, successfully implementing sim-to-real. We demonstrate our model on SDUQuad-48, and experiments show that the learned policy can achieve various dynamic motions with strong robustness.
DOI:10.1109/ROBIO55434.2022.10011921