Scholarlink

<p>Robust tactile texture recognition under realistic robotic interaction conditions remains challenging due to curvature-dependent contact variability and dynamic tactile excitation. This study proposes a Pressure-Conditioned Temporal Graph-Attention Network (PCT-GATNet) for multimodal tactile texture recognition using pressure-conditioned feature modulation, multi-scale temporal modeling, graph-based channel reasoning, and local temporal attention within a unified framework. The proposed method was evaluated on two public tactile benchmarks involving uneven curved-surface interaction and varying sliding velocities. Comparative experiments against CNN, TCN, and CNN–Transformer baselines were conducted under unified preprocessing and cross-validation protocols across temporal windows of 128, 256, and 512 samples. PCT-GATNet achieved 97.06% accuracy on the uneven-surface dataset under 10-fold cross-validation and up to 98% accuracy on the dynamic texture dataset, consistently outperforming the baseline architectures. Additional statistical analysis and quantitative explainability consistency evaluation demonstrated improved robustness, stable temporal representation learning, and physically meaningful tactile feature attribution. The findings demonstrate that pressure-conditioned graph-attention temporal modeling improves both performance and interpretability for tactile perception under realistic robotic interaction 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

Pressure-conditioned temporal graph-attention network for robotic tactile perception on uneven surfaces and varying velocities

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月26日

查看全文
  • Abdullah Alharthi

摘要

Robust tactile texture recognition under realistic robotic interaction conditions remains challenging due to curvature-dependent contact variability and dynamic tactile excitation. This study proposes a Pressure-Conditioned Temporal Graph-Attention Network (PCT-GATNet) for multimodal tactile texture recognition using pressure-conditioned feature modulation, multi-scale temporal modeling, graph-based channel reasoning, and local temporal attention within a unified framework. The proposed method was evaluated on two public tactile benchmarks involving uneven curved-surface interaction and varying sliding velocities. Comparative experiments against CNN, TCN, and CNN–Transformer baselines were conducted under unified preprocessing and cross-validation protocols across temporal windows of 128, 256, and 512 samples. PCT-GATNet achieved 97.06% accuracy on the uneven-surface dataset under 10-fold cross-validation and up to 98% accuracy on the dynamic texture dataset, consistently outperforming the baseline architectures. Additional statistical analysis and quantitative explainability consistency evaluation demonstrated improved robustness, stable temporal representation learning, and physically meaningful tactile feature attribution. The findings demonstrate that pressure-conditioned graph-attention temporal modeling improves both performance and interpretability for tactile perception under realistic robotic interaction conditions.