Scholarlink

<p>This paper presents a Simulation-to-Reality (Sim2Real) framework for robotic pick-and-place applications in unstructured and dynamic environments using deep reinforcement learning (DRL). A control policy is developed entirely in simulation via the proximal policy optimization (PPO) algorithm, utilizing the Isaac Lab and Isaac Sim platforms to ensure safe and efficient exploration. To prioritize real-time responsiveness and minimize computational overhead, a lightweight color-based detection module is employed to estimate 3D target positions from RGB-D images. The framework leverages a decoupled architecture where policy inference is separated from physical execution, thereby enhancing robustness against sensor noise and domain-specific variability. Experimental validation across diverse scenarios demonstrates an average success rate exceeding 85% in real-world applications. While achieving a 100% success rate in fixed static settings, the system maintains a high success rate of 93.3% in dynamic environments through real-time trajectory regeneration based on continuous object tracking. However, performance degradation to 60.0% in complex multi-object scenarios highlights persistent challenges in 3D depth estimation and unmodeled physical interactions during the grasping phase. These findings confirm the feasibility of zero-shot transfer for reaching tasks and provide a foundation for future advancements in adaptive robotic manipulation for collaborative human-robot environments.</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

Sim-to-Real Transfer of Deep Reinforcement Learning for Robotic Pick-and-Place Tasks

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Minseo Kim,
  • Do-Gyeong Yuk,
  • Jihyun Lee,
  • Jung Woo Sohn

摘要

This paper presents a Simulation-to-Reality (Sim2Real) framework for robotic pick-and-place applications in unstructured and dynamic environments using deep reinforcement learning (DRL). A control policy is developed entirely in simulation via the proximal policy optimization (PPO) algorithm, utilizing the Isaac Lab and Isaac Sim platforms to ensure safe and efficient exploration. To prioritize real-time responsiveness and minimize computational overhead, a lightweight color-based detection module is employed to estimate 3D target positions from RGB-D images. The framework leverages a decoupled architecture where policy inference is separated from physical execution, thereby enhancing robustness against sensor noise and domain-specific variability. Experimental validation across diverse scenarios demonstrates an average success rate exceeding 85% in real-world applications. While achieving a 100% success rate in fixed static settings, the system maintains a high success rate of 93.3% in dynamic environments through real-time trajectory regeneration based on continuous object tracking. However, performance degradation to 60.0% in complex multi-object scenarios highlights persistent challenges in 3D depth estimation and unmodeled physical interactions during the grasping phase. These findings confirm the feasibility of zero-shot transfer for reaching tasks and provide a foundation for future advancements in adaptive robotic manipulation for collaborative human-robot environments.