Scholarlink

This study utilizes Isaac Gym’s GPU capabilities and a Recurrent Neural Network (RNN) [1] trained with the Proximal Policy Optimization (PPO) [3] algorithm to improve quadruped robots’ mobility in complex environments. A precise URDF model was used to ensure accurate training parameters, and an RNN processed proprioceptive [4] data for action output. The large-scale training of 4,192 robots on this platform significantly enhanced learning efficiency, enabling effective terrain navigation and robust performance. Integrating RNNs enables robots to adapt to dynamic environments through informed decision-making. The framework enhances agility and stability while optimizing training with PPO, significantly improving quadruped robot performance.

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

Quadruped Robot System Based on Reinforcement Learning and Parallel Simulation Training

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xiaobo Wang,
  • Zhencheng Ye,
  • Xufeng Ling,
  • Lei Wu

摘要

This study utilizes Isaac Gym’s GPU capabilities and a Recurrent Neural Network (RNN) [1] trained with the Proximal Policy Optimization (PPO) [3] algorithm to improve quadruped robots’ mobility in complex environments. A precise URDF model was used to ensure accurate training parameters, and an RNN processed proprioceptive [4] data for action output. The large-scale training of 4,192 robots on this platform significantly enhanced learning efficiency, enabling effective terrain navigation and robust performance. Integrating RNNs enables robots to adapt to dynamic environments through informed decision-making. The framework enhances agility and stability while optimizing training with PPO, significantly improving quadruped robot performance.