Scholarlink

<p>Intelligent robots or autonomous service robots represent some of the most proverbial landmarks of the Industry 4.0 era. Compared to traditional approaches, achieving this goal requires robots to perceive their environment, reason about it, and act with human-like cognitive capabilities. Therefore, a perception-based framework utilizing visual language model-reinforcement learning and incorporating synergistic push–grasp actions (VLM-RLPGS) for industrial manipulator has been proposed. In this paper, our approach goes as follows: (i) A precise description of the target framework and design objectives is established; (ii) evaluation of the autonomous execution mechanism for computing correct actions is conducted; and (iii) expert learning process, e.g., geometric analysis from input images, reward function design, and training with sample datasets, is discussed. To verify the feasibility of and efficiency in the proposed approach, several numerical simulations according to our framework design are conducted. Consequently, an experimental platform, consisting of a Universal Robot UR5, a two-finger gripper, a stereo camera, and sample objects, is then launched. In accordance with the results of both simulations and experiments, the effectiveness, applicability, and potential scalability of our approach to different areas can be precisely substantiated.</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

VLM-RLPGS: A Cognitive Framework Using Vision–Language Model and Reinforcement Learning for Push–Grasp Synergy

  • 期刊论文

  • 发表日期：

    2026年2月22日

查看全文
  • Thanh Phuong Nguyen,
  • Hung Nguyen,
  • Song Hung Nguyen,
  • Ha Quang Thinh Ngo

摘要

Intelligent robots or autonomous service robots represent some of the most proverbial landmarks of the Industry 4.0 era. Compared to traditional approaches, achieving this goal requires robots to perceive their environment, reason about it, and act with human-like cognitive capabilities. Therefore, a perception-based framework utilizing visual language model-reinforcement learning and incorporating synergistic push–grasp actions (VLM-RLPGS) for industrial manipulator has been proposed. In this paper, our approach goes as follows: (i) A precise description of the target framework and design objectives is established; (ii) evaluation of the autonomous execution mechanism for computing correct actions is conducted; and (iii) expert learning process, e.g., geometric analysis from input images, reward function design, and training with sample datasets, is discussed. To verify the feasibility of and efficiency in the proposed approach, several numerical simulations according to our framework design are conducted. Consequently, an experimental platform, consisting of a Universal Robot UR5, a two-finger gripper, a stereo camera, and sample objects, is then launched. In accordance with the results of both simulations and experiments, the effectiveness, applicability, and potential scalability of our approach to different areas can be precisely substantiated.