Scholarlink

<p>The perception in most existing vision-based reinforcement learning (RL) models for robotic manipulation relies heavily on static third-person or hand-mounted first-person cameras. In scenarios with occlusions and limited maneuvering space, these carefully positioned cameras often struggle to provide effective visual observations during manipulation. Taking inspiration from human capabilities, we introduce a novel RL-based dual-arm active visual-guided manipulation model (DAVMM), which simultaneously infers “eye” actions and “hand” actions for two separate robotic arms (referred to as the vision-arm and the worker-arm) based on current observations, empowering the robot with the ability to actively perceive and interact with its environment. To handle the extensive redundant observation-action space, we propose a decouplable target-centric reward paradigm to offer stable guidance for the training process. For making fine-grained manipulation action decisions, alongside a global scene image encoder, we utilize an independent encoder to extract local target texture features, enabling the simultaneous acquisition of both global and detailed local information. Additionally, we employ residual-RL and curriculum learning techniques to further enhance our model’s sample efficiency and training stability. We conducted comparative experiments and analyses of DAVMM against a set of strong baselines on three occluded and narrow-space manipulation tasks. DAVMM notably improves the success rates across all manipulation tasks and showcases rapid learning capabilities.</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

One model, two skills: active vision and action learning model for robotic manipulation

  • 期刊论文

  • 发表日期：

    2025年3月21日

查看全文
  • Guokang Wang,
  • Yanhong Liu,
  • Huaping Liu

摘要

The perception in most existing vision-based reinforcement learning (RL) models for robotic manipulation relies heavily on static third-person or hand-mounted first-person cameras. In scenarios with occlusions and limited maneuvering space, these carefully positioned cameras often struggle to provide effective visual observations during manipulation. Taking inspiration from human capabilities, we introduce a novel RL-based dual-arm active visual-guided manipulation model (DAVMM), which simultaneously infers “eye” actions and “hand” actions for two separate robotic arms (referred to as the vision-arm and the worker-arm) based on current observations, empowering the robot with the ability to actively perceive and interact with its environment. To handle the extensive redundant observation-action space, we propose a decouplable target-centric reward paradigm to offer stable guidance for the training process. For making fine-grained manipulation action decisions, alongside a global scene image encoder, we utilize an independent encoder to extract local target texture features, enabling the simultaneous acquisition of both global and detailed local information. Additionally, we employ residual-RL and curriculum learning techniques to further enhance our model’s sample efficiency and training stability. We conducted comparative experiments and analyses of DAVMM against a set of strong baselines on three occluded and narrow-space manipulation tasks. DAVMM notably improves the success rates across all manipulation tasks and showcases rapid learning capabilities.