Scholarlink

<p>This paper introduces a safety-focused, learning-based approach for object manipulation tasks in dynamic environments where human–robot collaboration is essential, ensuring adaptive and safety-aware operations. Our methodology employs simulation data to train and test policies using two different simulation frameworks: Handoversim, a recently developed benchmark for human-to-robot object handovers, and our custom simulation framework. Both frameworks are integrated with the OpenAI Safety Gym library, which supports three mobile agent types learning goal-directed tasks under various safety constraints but did not initially include a robotic arm. We have integrated a new model, the Panda robotic arm, and validated its applicability within Safety Gym. Our focus is on accurate and secure object manipulation, particularly in human–robot interaction scenarios, within the custom simulation framework end-to-end policy achieved successfully. Handoversim, while advanced, struggled with object handling due to inherent challenges. This led to established baselines switching to hand-coded policies for grasping instead of relying on end-to-end policy learning. We compared a model-free proximal policy optimization (PPO) baseline with its constrained variant (cPPO) within a reinforcement learning framework. To accelerate learning while maintaining safety compliance, we introduce "Grasp Mechanics", an innovative tactile feedback framework that reduces dependence on complex vision-based systems. This approach emulates human-like object interaction through proprioceptive feedback, enabling precise and secure manipulation without requiring detailed prior object knowledge. Preliminary experiments indicate that although cPPO requires longer training periods, it achieves policy performance comparable to the baseline PPO, while significantly outperforming it in terms of safety compliance. This paper highlights the potential of integrating advanced reinforcement learning techniques with robust safety mechanisms, as provided by Safety Gym, to elevate the capabilities of robotic systems in safely handling safety-critical manipulation tasks, crucial for real-world applications. Our findings lay the groundwork for future advancements in safe autonomous robotics, emphasizing the critical role of integrating safety considerations from the ground up.</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

Reinforcement learning for precision grasping and safety-critical coordination in a robotic arm

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Fawad Khan,
  • Wei Feng,
  • Tianlun Huang,
  • Zhiyong Wang,
  • Xiao Liu,
  • Shahid Asad Ali,
  • Weijun Wang

摘要

This paper introduces a safety-focused, learning-based approach for object manipulation tasks in dynamic environments where human–robot collaboration is essential, ensuring adaptive and safety-aware operations. Our methodology employs simulation data to train and test policies using two different simulation frameworks: Handoversim, a recently developed benchmark for human-to-robot object handovers, and our custom simulation framework. Both frameworks are integrated with the OpenAI Safety Gym library, which supports three mobile agent types learning goal-directed tasks under various safety constraints but did not initially include a robotic arm. We have integrated a new model, the Panda robotic arm, and validated its applicability within Safety Gym. Our focus is on accurate and secure object manipulation, particularly in human–robot interaction scenarios, within the custom simulation framework end-to-end policy achieved successfully. Handoversim, while advanced, struggled with object handling due to inherent challenges. This led to established baselines switching to hand-coded policies for grasping instead of relying on end-to-end policy learning. We compared a model-free proximal policy optimization (PPO) baseline with its constrained variant (cPPO) within a reinforcement learning framework. To accelerate learning while maintaining safety compliance, we introduce "Grasp Mechanics", an innovative tactile feedback framework that reduces dependence on complex vision-based systems. This approach emulates human-like object interaction through proprioceptive feedback, enabling precise and secure manipulation without requiring detailed prior object knowledge. Preliminary experiments indicate that although cPPO requires longer training periods, it achieves policy performance comparable to the baseline PPO, while significantly outperforming it in terms of safety compliance. This paper highlights the potential of integrating advanced reinforcement learning techniques with robust safety mechanisms, as provided by Safety Gym, to elevate the capabilities of robotic systems in safely handling safety-critical manipulation tasks, crucial for real-world applications. Our findings lay the groundwork for future advancements in safe autonomous robotics, emphasizing the critical role of integrating safety considerations from the ground up.