Scholarlink

<p>In the modern era of Deep Learning, network parameters are crucial for model efficiency but come with drawbacks like high computational demands and memory requirements, making them less suitable for real-time intelligent robot grasping tasks. Current research in vision-based robotics aims to enhance model efficiency by introducing sparsity without sacrificing accuracy in robot grasp pose detection. Here we introduce two streamlined neural network architectures: Sparse-Generative Residual ConvNet (GRConvNet) and Sparse-Generative Inception Neural Network (GINNet). These architectures leverage sparsity in robotic grasping by integrating the Edge-PopUp algorithm to generate high-quality grasp poses in real-time at every pixel location, enabling effective manipulation of unfamiliar objects by robots. Our models underwent extensive training on two benchmark datasets, the Cornell Grasping Dataset (CGD) and the Jacquard Grasping Dataset (JGD). Remarkably, our models performed comparably to state-of-the-art methods. Sparse-GRConvNet achieved an accuracy of 97.75% with only 10% of the network parameters of GR-ConvNet on CGD and an accuracy of 85.77% with 30% of the parameters of GR-ConvNet on JGD. Sparse-GINNet achieved an accuracy of 97.75% with 50% of the weight of GI-NNet on CGD and an accuracy of 81.11% with 10% of the weight of GI-NNet on JGD. The novel method emphasizes more on finding the optimum network architecture and does not perform conventional weight training. Rather our algorithm tries to find the high-scored edges of the network for finding a potential subnetwork having sparsity. Extensive experiments using the Anukul (Baxter) hardware cobot validated the performance of our proposed models.</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

Vision-based intelligent robot grasping using sparse neural network

  • 期刊论文

  • 发表日期：

    2025年2月21日

查看全文
  • Vandana Kushwaha,
  • Priya Shukla,
  • G. C. Nandi

摘要

In the modern era of Deep Learning, network parameters are crucial for model efficiency but come with drawbacks like high computational demands and memory requirements, making them less suitable for real-time intelligent robot grasping tasks. Current research in vision-based robotics aims to enhance model efficiency by introducing sparsity without sacrificing accuracy in robot grasp pose detection. Here we introduce two streamlined neural network architectures: Sparse-Generative Residual ConvNet (GRConvNet) and Sparse-Generative Inception Neural Network (GINNet). These architectures leverage sparsity in robotic grasping by integrating the Edge-PopUp algorithm to generate high-quality grasp poses in real-time at every pixel location, enabling effective manipulation of unfamiliar objects by robots. Our models underwent extensive training on two benchmark datasets, the Cornell Grasping Dataset (CGD) and the Jacquard Grasping Dataset (JGD). Remarkably, our models performed comparably to state-of-the-art methods. Sparse-GRConvNet achieved an accuracy of 97.75% with only 10% of the network parameters of GR-ConvNet on CGD and an accuracy of 85.77% with 30% of the parameters of GR-ConvNet on JGD. Sparse-GINNet achieved an accuracy of 97.75% with 50% of the weight of GI-NNet on CGD and an accuracy of 81.11% with 10% of the weight of GI-NNet on JGD. The novel method emphasizes more on finding the optimum network architecture and does not perform conventional weight training. Rather our algorithm tries to find the high-scored edges of the network for finding a potential subnetwork having sparsity. Extensive experiments using the Anukul (Baxter) hardware cobot validated the performance of our proposed models.