Trajectory Tracking Control for Robotic Manipulator Based on Soft Actor-Critic and Generative Adversarial Imitation...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Robotic manipulators face challenges in tracking control due to saturation constraints and random disturbances.
• Existing methods often require complex dynamic and kinematic models, which can be difficult to obtain.
• Training instability is a common issue in reinforcement learning for robotic control.

Purpose of the Study:

• To propose a novel deep reinforcement learning (DRL) approach for robotic manipulator tracking control.
• To address saturation constraints and random disturbances without explicit model learning.
• To enhance training stability and control policy performance.

Main Methods:

• A Soft Actor-Critic (SAC) and Long Short-Term Memory (LSTM) combination (SAC-LSTM) was developed to improve training stability by capturing joint position trends.
• Generative Adversarial Imitation Learning (GAIL) was employed, using the SAC-LSTM policy as expert data to learn a superior control policy.
• The integrated SAC-LSTM-GAIL (SL-GAIL) algorithm was designed to learn directly from stable expert data, avoiding environment exploration.

Main Results:

• The SL-GAIL algorithm effectively accomplished robotic manipulator end-effector tracking tasks in simulations.
• The proposed method demonstrated superior stability compared to other algorithms, particularly in environments with interference.
• The approach successfully managed torque and joint angle saturation constraints.

Conclusions:

• The SL-GAIL algorithm provides an effective and stable solution for robotic manipulator tracking control under challenging conditions.
• This model-free approach simplifies the control strategy learning process by leveraging expert data.
• The findings highlight the potential of combining LSTM, SAC, and GAIL for advanced robotic control applications.