Scholarlink

In recent years, humanoid robots have demonstrated significant potential in completing complex tasks in dynamic environments. Gate-passing is a challenging task that requires precise motion control and decision-making capabilities. This paper proposes a humanoid robots gate-passing method based on the Soft Actor-Critic (SAC) algorithm combined with Generative Adversarial Networks (GAN). By incorporating a GAN module to generate adversarial samples, the SAC algorithm receives additional feedback, enhancing the stability and efficiency of policy generation. Experiments conducted on the Webots simulation platform using the DarwinOP2 humanoid robots demonstrate that the GAN-SAC method significantly accelerates convergence and improves task success rates compared to traditional SAC. This study addresses the challenges of unstable policy generation and inefficient exploration in complex dynamic environments, while significantly enhancing the robot’s task adaptability in random scenarios.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Humanoid Robots Gate-Passing Method Based on SAC Algorithm with Generative Adversarial Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhiyuan Chen,
  • Jianhua Dong,
  • Guodong Zhao,
  • Mingshuo Liu,
  • Shuaiqi Zhang

摘要

In recent years, humanoid robots have demonstrated significant potential in completing complex tasks in dynamic environments. Gate-passing is a challenging task that requires precise motion control and decision-making capabilities. This paper proposes a humanoid robots gate-passing method based on the Soft Actor-Critic (SAC) algorithm combined with Generative Adversarial Networks (GAN). By incorporating a GAN module to generate adversarial samples, the SAC algorithm receives additional feedback, enhancing the stability and efficiency of policy generation. Experiments conducted on the Webots simulation platform using the DarwinOP2 humanoid robots demonstrate that the GAN-SAC method significantly accelerates convergence and improves task success rates compared to traditional SAC. This study addresses the challenges of unstable policy generation and inefficient exploration in complex dynamic environments, while significantly enhancing the robot’s task adaptability in random scenarios.