Reinforcement Learning Tracking Control for Robotic Manipulator With Kernel-Based Dynamic Model.

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Reinforcement learning (RL) is effective for robot control but faces challenges in continuous tasks.
• Existing methods often require knowledge of the robot's dynamic model, limiting applicability.

Purpose of the Study:

• To propose a novel kernel-based dynamic model for reinforcement learning in robotic manipulators.
• To develop an efficient RL tracking controller that does not require a pre-defined dynamic model.
• To enhance the learning speed and performance of robotic control tasks.

Main Methods:

• A kernel-based dynamic model is developed for RL, eliminating the need to learn the manipulator's dynamics.
• A new tuple is formed using kernel function sampling to define the robotic RL control problem.
• A reward function tailored for tracking control is defined to accelerate learning, coupled with a critic system for policy evaluation.

Main Results:

• The proposed method effectively fulfills robotic tracking tasks.
• Achieves comparable or superior tracking performance compared to other learning algorithms.
• Demonstrates significantly reduced force/torque input requirements, highlighting efficiency.

Conclusions:

• The kernel-based dynamic model offers an effective and efficient solution for robotic tracking control using RL.
• The approach successfully addresses challenges in continuous control tasks without explicit dynamic model learning.
• Validated effectiveness and efficiency through simulation results, showing reduced energy consumption.