Scholarlink

This paper presents a reinforcement learning (RL) approach to automate bin-picking tasks for small and medium-sized enterprises (SMEs), addressing the need for cost-effective and flexible automation. Using a UR5 collaborative robot (cobot) in a CoppeliaSim environment, the method employs a Deep Q-Network (DQN) with ResNet-based feature extraction to predict grasp success probabilities. A key innovation is the early termination of exploration once the exploitation accuracy exceeds 85%, thereby enhancing sample efficiency and reducing training time. The model achieves a final total reward of 278 out of 300 episodes, demonstrating significant improvement over baseline methods. By enabling autonomous reprogramming, this RL framework reduces dependence on specialized robotics expertise, offering SMEs a practical and adaptable bin-picking solution.

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

Accelerating SME Robotics: Reinforcement Learning for Efficient Bin-Picking

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Philippe Juhel

摘要

This paper presents a reinforcement learning (RL) approach to automate bin-picking tasks for small and medium-sized enterprises (SMEs), addressing the need for cost-effective and flexible automation. Using a UR5 collaborative robot (cobot) in a CoppeliaSim environment, the method employs a Deep Q-Network (DQN) with ResNet-based feature extraction to predict grasp success probabilities. A key innovation is the early termination of exploration once the exploitation accuracy exceeds 85%, thereby enhancing sample efficiency and reducing training time. The model achieves a final total reward of 278 out of 300 episodes, demonstrating significant improvement over baseline methods. By enabling autonomous reprogramming, this RL framework reduces dependence on specialized robotics expertise, offering SMEs a practical and adaptable bin-picking solution.