An Efficient Robotic Pushing and Grasping Method in Cluttered Scene.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Pushing and grasping (PG) are essential for intelligent robots.
• Current PG methods are either fast but inaccurate (single-stage) or accurate but slow (multistage).

Purpose of the Study:

• To propose a novel end-to-end method, EPGNet, that achieves both high accuracy and efficiency in robotic PG tasks.
• To optimize performance with fewer parameters using EfficientNet-B0 backbone.

Main Methods:

• Developed EPGNet, an end-to-end network with a cross-fusion module for fusing local and global features.
• Utilized a Q-learning framework for simultaneous training of two PG action prediction branches.
• Created a unique PG dataset and trained a YOLACT network for object detection and segmentation to bridge the sim-to-real gap.

Main Results:

• EPGNet demonstrated superior performance compared to single-stage PG methods.
• EPGNet achieved competitive results against multistage methods.
• The proposed method effectively handles objects of varying sizes in different scenes with fewer parameters.

Conclusions:

• EPGNet offers a balanced solution for robotic pushing and grasping, achieving high accuracy and efficiency.
• The novel cross-fusion module and training strategy contribute to robust performance in complex robotic tasks.