Robot grasping method optimization using improved deep deterministic policy gradient algorithm of deep reinforcement...

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robot grasping is crucial but challenged by inefficiencies in traditional methods when dealing with unknown disturbances.
• Existing deep learning and coordinate positioning methods struggle with grasping accuracy when target objects move or are disturbed.

Purpose of the Study:

• To enhance robot grasping efficiency and robustness against unknown disturbances.
• To develop an autonomous learning framework for robot grasping tasks.

Main Methods:

• Proposed a Gaussian parameter Deep Deterministic Policy Gradient (Gaussian-DDPG) algorithm integrated with an Importance-Weighted Autoencoder (IWAE).
• Utilized IWAE to compress high-dimensional visual input into a hidden space for the reinforcement learning network.
• Enhanced the DDPG algorithm with Gaussian parameters for improved exploration and dynamic workspace adaptation.

Main Results:

• The proposed Gaussian-DDPG with IWAE enables autonomous learning of grasping tasks.
• Achieved improved grasping efficiency and accuracy, particularly in disturbed situations with moving objects.
• Optimized manipulator torque control using visual information deviation for precise grasping.

Conclusions:

• The novel Gaussian-DDPG algorithm effectively addresses limitations in traditional robot grasping methods.
• The integration of IWAE and Gaussian parameters allows robots to learn and adapt to complex grasping scenarios.
• This approach significantly improves the robot's ability to handle unknown disturbances and achieve accurate manipulation.