Scholarlink

<p>A foundational framework for diverse robot motion generation is proposed and its effectiveness in creating appealing robots is verified in this paper. The framework uses a two-step approach: train a large language model (LLM) and use it to control the robot’s behavior. During the learning phase, the training dataset is generated by extracting 2D keypoints from cat videos, converting these 2D keypoints into 3D keypoints, and calculating joint angles using inverse kinematics based on joint positions. These data, comprising joints and their positions, were transformed into a language-like format, termed “motion language.” These converted data are then used to train the gMLP, a large language model. The gMLP showed its lowest validation loss after 2 epochs, so the model trained up to this point is used for motion generation for over 30&#xa0;min. These motion data generated by trained gMLP were analyzed by clustering method. The results showed that the analysis of 5-s intervals revealed five distinct movement patterns. Similarly, in 10-s intervals, three distinct movement patterns were observed.</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

Clustering analysis of long-term robot motion patterns generated by cat motion trained gMLP

  • 期刊论文

  • 发表日期：

    2025年10月22日

查看全文
  • Koki Sato,
  • Ryo Kobayashi,
  • Sho Yamauchi,
  • Keiji Suzuki,
  • Sho’ji Suzuki

摘要

A foundational framework for diverse robot motion generation is proposed and its effectiveness in creating appealing robots is verified in this paper. The framework uses a two-step approach: train a large language model (LLM) and use it to control the robot’s behavior. During the learning phase, the training dataset is generated by extracting 2D keypoints from cat videos, converting these 2D keypoints into 3D keypoints, and calculating joint angles using inverse kinematics based on joint positions. These data, comprising joints and their positions, were transformed into a language-like format, termed “motion language.” These converted data are then used to train the gMLP, a large language model. The gMLP showed its lowest validation loss after 2 epochs, so the model trained up to this point is used for motion generation for over 30 min. These motion data generated by trained gMLP were analyzed by clustering method. The results showed that the analysis of 5-s intervals revealed five distinct movement patterns. Similarly, in 10-s intervals, three distinct movement patterns were observed.