Scholarlink

<p>Quadruped robots have inherent advantages in unstructured terrain, but agile motion control remains challenging. These challenges include insufficient 3D terrain perception and a significant visual domain shift between simulation and reality. Moreover, integrating diverse motion skills remains difficult. This paper proposes an end-to-end motion policy framework to address these issues by integrating precise 3D perception and sensor characteristic modeling. A 3D composite scanning mechanism with three vertical layers and one horizontal layer is designed. It efficiently perceives 3D structures such as overhanging obstacles, without relying on elevation maps or world models. Six types of physical noise from depth cameras are systematically modeled. With simulated image degradation and real-world data homogenization, the distribution of real depth images is accurately reproduced. Based on the teacher-student framework and Dataset Aggregation (DAgger) knowledge distillation, diverse motion skills and zero-shot transfer are integrated end-to-end. Experiments show that the framework enables quadruped robots to climb vertical obstacles, cross gaps, crawl under obstacles, and move stably on sloped grasslands.</p>

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

Agile Motion Control of Quadruped Robots on Unstructured Terrain Based on 3D Composite Perception and Depth Camera Characteristic Modeling

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Yingzhong Tian,
  • Chuan Liu,
  • Tao Jin,
  • Long Li,
  • Jieyu Wang

摘要

Quadruped robots have inherent advantages in unstructured terrain, but agile motion control remains challenging. These challenges include insufficient 3D terrain perception and a significant visual domain shift between simulation and reality. Moreover, integrating diverse motion skills remains difficult. This paper proposes an end-to-end motion policy framework to address these issues by integrating precise 3D perception and sensor characteristic modeling. A 3D composite scanning mechanism with three vertical layers and one horizontal layer is designed. It efficiently perceives 3D structures such as overhanging obstacles, without relying on elevation maps or world models. Six types of physical noise from depth cameras are systematically modeled. With simulated image degradation and real-world data homogenization, the distribution of real depth images is accurately reproduced. Based on the teacher-student framework and Dataset Aggregation (DAgger) knowledge distillation, diverse motion skills and zero-shot transfer are integrated end-to-end. Experiments show that the framework enables quadruped robots to climb vertical obstacles, cross gaps, crawl under obstacles, and move stably on sloped grasslands.