Scholarlink

<p>Vision-based robotic grasping stands as a core technology in interactive tasks. While existing methods demonstrate exceptional performance on annotated datasets, practical grasping accuracy significantly deteriorates under environmental variations, illumination fluctuations, and viewpoint disparities. To address this challenge, we propose a domain-adaptive grasping network that integrates the Mamba architecture with wavelet pooling and pinwheel-shaped convolution. The framework innovatively employs a dual-stream processing mechanism: For RGB images rich in semantic information, we leverage the Mamba model’s long-range dependency modeling capabilities for feature extraction; For depth data, we design a pinwheel-shaped convolution with large receptive fields combined with wavelet pooling to capture geometric features, subsequently constructing a baseline model through feature fusion. In terms of training strategy, we introduce a CORAL domain discrepancy alignment loss to jointly optimize the network using Cornell annotated datasets and unlabeled real-world data, achieving cross-domain feature alignment through backpropagation. Experimental results reveal that the baseline model achieves 99.25% accuracy on the Cornell dataset but only 71.43% in real-world scenarios. After domain adaptation implementation, Cornell accuracy slightly decreases to 96.26%, while real-world performance significantly improves to 90.48%, demonstrating the method’s effectiveness in enhancing generalization capabilities for unseen objects. This study provides a novel cross-domain solution for robotic grasping systems. Code and data collection was published on <a href="https://github.com/BaiyangWang256/DW-MambaNet/tree/main">https://github.com/BaiyangWang256/DW-MambaNet/tree/main</a>.</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

Mamba and wavelet-enhanced dual-modal domain adaptation for grasp detection

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Baiyang Wang,
  • Ming Fang,
  • Na Wang,
  • Fang Li,
  • Yang Li,
  • Xujian Wang,
  • Hongjun Wang,
  • Hua Li

摘要

Vision-based robotic grasping stands as a core technology in interactive tasks. While existing methods demonstrate exceptional performance on annotated datasets, practical grasping accuracy significantly deteriorates under environmental variations, illumination fluctuations, and viewpoint disparities. To address this challenge, we propose a domain-adaptive grasping network that integrates the Mamba architecture with wavelet pooling and pinwheel-shaped convolution. The framework innovatively employs a dual-stream processing mechanism: For RGB images rich in semantic information, we leverage the Mamba model’s long-range dependency modeling capabilities for feature extraction; For depth data, we design a pinwheel-shaped convolution with large receptive fields combined with wavelet pooling to capture geometric features, subsequently constructing a baseline model through feature fusion. In terms of training strategy, we introduce a CORAL domain discrepancy alignment loss to jointly optimize the network using Cornell annotated datasets and unlabeled real-world data, achieving cross-domain feature alignment through backpropagation. Experimental results reveal that the baseline model achieves 99.25% accuracy on the Cornell dataset but only 71.43% in real-world scenarios. After domain adaptation implementation, Cornell accuracy slightly decreases to 96.26%, while real-world performance significantly improves to 90.48%, demonstrating the method’s effectiveness in enhancing generalization capabilities for unseen objects. This study provides a novel cross-domain solution for robotic grasping systems. Code and data collection was published on https://github.com/BaiyangWang256/DW-MambaNet/tree/main.