Research on Small Sample Multi-Target Grasping Technology Based on Transfer Learning.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Robot multi-target grasping detection is crucial for advanced automation.
• Existing methods face challenges in accurately identifying and localizing multiple objects for grasping.
• The integration of attention mechanisms and spatial pyramid pooling offers potential for improved feature extraction.

Purpose of the Study:

• To propose and validate a new model, CBAM-ASPP-SqueezeNet, for enhanced robot multi-target grasping detection.
• To leverage transfer learning for efficient model training on a custom multi-target grasping dataset.
• To improve the SqueezeNet architecture using channel and spatial attention with atrous spatial pyramid pooling.

Main Methods:

• Development and expansion of a multi-target grasping dataset.
• Application of transfer learning for pre-training on a single-target dataset.
• Optimization of the SqueezeNet model with the Convolutional Block Attention Module (CBAM) and Atrous Spatial Pyramid Pooling (ASPP).
• Implementation of attention mechanisms for feature map weighting in channel and spatial dimensions.
• Utilizing atrous convolutions with varying rates to expand the receptive field and capture multi-scale features.

Main Results:

• The CBAM-ASPP-SqueezeNet model demonstrated convergence after 20 epochs of training with transfer learning.
• A significant grabbing success rate of 93% was achieved in physical experiments using Kinova and SIASUN robotic arms.
• The proposed attention and ASPP modules effectively improved feature representation for grasping tasks.

Conclusions:

• The CBAM-ASPP-SqueezeNet model is effective for robot multi-target grasping detection.
• Transfer learning accelerates convergence and improves performance on custom datasets.
• The integration of attention mechanisms and ASPP enhances the capability of robotic systems in complex manipulation tasks.