Scholarlink

<p>Functional grasping of tools with initial linear/angular velocities in zero-gravity environments presents significant challenges. Unlike static tabletop grasping, where object motion can be restricted through surface contact, the absence of gravitational constraints allows minor collisions to propel tools away, leading to frequent grasping failures. Additionally, robotic hands with fewer degrees of freedom compared to human hands struggle to replicate human-like functional grasping postures. This work addresses these challenges through human-inspired heuristic reward design, employing reinforcement learning to enable dexterous hands to autonomously explore feasible grasping postures in simulation. This paper systematically summarizes the end-effector approach patterns for zero-gravity floating grasping and develops corresponding reward functions. Using the LSTM-PPO (a hybrid architecture combining long short-term memory (LSTM) networks with proximal policy optimization (PPO)) method within the Isaac Gym graphics processing unit (GPU)-accelerated parallel simulation environment, functional grasping of floating tools was successfully achieved, attaining an average success rate of 56.2%, outperforming several baseline approaches, including inverse kinematics and an multi-layer perceptron (MLP)-only PPO policy. Key contributions include a domain-specific reward-shaping framework and actionable insights into dynamic tool manipulation in microgravity conditions.</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

Functional grasping of floating tools in zero-gravity space using reinforcement learning for dexterous robotic hands

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Yifan Wang,
  • Houcheng Li,
  • Mingyang Li,
  • Shuanfeng Xu,
  • Nailong Liu,
  • Long Cheng

摘要

Functional grasping of tools with initial linear/angular velocities in zero-gravity environments presents significant challenges. Unlike static tabletop grasping, where object motion can be restricted through surface contact, the absence of gravitational constraints allows minor collisions to propel tools away, leading to frequent grasping failures. Additionally, robotic hands with fewer degrees of freedom compared to human hands struggle to replicate human-like functional grasping postures. This work addresses these challenges through human-inspired heuristic reward design, employing reinforcement learning to enable dexterous hands to autonomously explore feasible grasping postures in simulation. This paper systematically summarizes the end-effector approach patterns for zero-gravity floating grasping and develops corresponding reward functions. Using the LSTM-PPO (a hybrid architecture combining long short-term memory (LSTM) networks with proximal policy optimization (PPO)) method within the Isaac Gym graphics processing unit (GPU)-accelerated parallel simulation environment, functional grasping of floating tools was successfully achieved, attaining an average success rate of 56.2%, outperforming several baseline approaches, including inverse kinematics and an multi-layer perceptron (MLP)-only PPO policy. Key contributions include a domain-specific reward-shaping framework and actionable insights into dynamic tool manipulation in microgravity conditions.