Reinforcement Learning with Task Decomposition and Task-Specific Reward System for Automation of High-Level Tasks

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Complex high-level robotic tasks present significant challenges for traditional end-to-end learning methods.
• Task decomposition and tailored reward systems are crucial for improving learning efficiency and success rates.

Purpose of the Study:

• To introduce a novel reinforcement learning method that combines task decomposition with a task-specific reward system.
• To enhance learning speed, success rates, and efficiency in executing complex robotic tasks.

Main Methods:

• Decomposition of complex tasks into simpler subtasks.
• Utilizing single joint and gripper actions for grasping and placing subtasks.
• Employing the Soft Actor-Critic (SAC) algorithm with a task-specific reward system for other subtasks.

Main Results:

• Achieved a 99.9% success rate for door opening.
• Attained a 95.25% success rate for block stacking.
• Demonstrated 80.8% and 90.9% success rates for square-nut and round-nut assembly, respectively.

Conclusions:

• The proposed reinforcement learning method effectively addresses complex robotic tasks through task decomposition and specialized rewards.
• This approach offers significant improvements in learning speed, success rates, and task execution efficiency compared to traditional methods.