Target recognition and grasping strategies for soft robotic manipulators in unstructured environments

Area of Science:

• Robotics
• Computer Vision
• Materials Science

Background:

• Robots struggle with grasping irregular and fragile objects in unstructured environments.
• Existing methods lack efficiency and accuracy for delicate manipulation tasks.

Purpose of the Study:

• To develop a soft robotic hand and an enhanced object detection algorithm for efficient and accurate grasping.
• To improve the performance of the You Only Look Once v5s (YOLOv5s) algorithm for real-time robotic applications.

Main Methods:

• Designed a rapid pneumatic network-based soft finger structure with a validated mathematical model for finger bending.
• Enhanced YOLOv5s by integrating Coordinate Attention (CA) and refining the Spatial Pyramid Pooling (SPP) module.
• Developed a soft robotic grasping experimental platform for precision and power grasping tests.

Main Results:

• The enhanced YOLOv5s-CA-SPP model showed improvements in mean average precision and recognition speed.
• Experimental analysis confirmed the effectiveness of the soft robotic hand and the improved detection algorithm.
• The system achieved an 82% success rate in grasping experiments with optimal posture.

Conclusions:

• The integrated soft robotic hand and enhanced YOLOv5s algorithm offer a viable solution for complex grasping tasks.
• The developed system demonstrates significant potential for applications in unstructured robotic manipulation.
• Further research can explore advanced control strategies and broader object recognition capabilities.