Eye-in-Hand Robotic Arm Gripping System Based on Machine Learning and State Delay Optimization.

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Existing robotic grasping systems often struggle with delicate tasks and time delays.
• State time delays in control systems are complex and less studied than input/output delays.
• Human-like grippers offer potential for improved dexterity.

Purpose of the Study:

• To develop an optimized control system for robotic grasping that accounts for state time delays.
• To evaluate the performance of a novel five-finger gripper compared to traditional grippers.
• To enhance machine learning network architecture for accurate grasp prediction using RGB-D images.

Main Methods:

• Modified an existing machine learning network architecture for grasp prediction.
• Utilized RGB-D images for object location prediction.
• Tested a five-finger gripper and a two-finger gripper on a 6-DOF robot arm, performing over 100 grasps.
• Investigated and proposed a control method for state time delays in robot manipulators.

Main Results:

• The five-finger gripper demonstrated superior performance in delicate tasks compared to two- or three-finger grippers.
• The modified network achieved high performance with limited training data and without posture evaluation.
• The integrated system, combining the network and the five-finger gripper, exceeded a 90% gripping success rate.
• The proposed control method effectively addressed state time delays.

Conclusions:

• The developed automated system significantly improves robotic grasping success rates.
• The five-finger gripper offers advanced dexterity for complex manipulation tasks.
• The modified machine learning approach provides an efficient solution for grasp prediction, even with limited data and system delays.