Koopman Operator-Based Knowledge-Guided Reinforcement Learning for Safe Human-Robot Interaction.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Task-constrained path generation for robotic manipulators is challenging.
• Leveraging human demonstrated trajectories offers a promising approach for robot learning.
• Existing methods may struggle to accurately capture human intent in complex tasks.

Purpose of the Study:

• To develop a novel framework for deep reinforcement learning (DRL) in robotic manipulator path generation.
• To design a versatile reward function using Koopman operator theory to model human intent from demonstrations.
• To enhance robot learning by incorporating human feedback when intent prediction is uncertain.

Main Methods:

• Developed a novel framework integrating DRL with human demonstrated trajectories.
• Utilized Koopman operator theory to build a human intent model from demonstrated paths.
• Created a trust domain for Koopman operator-based intent prediction, incorporating human feedback otherwise.
• Integrated the designed reward function into the Deep Q-Learning (DQN) framework, creating a modified DQN algorithm.

Main Results:

• Demonstrated the effectiveness of the proposed learning algorithm on a simulated robotic arm.
• Successfully learned paths for constrained end-effector motion.
• Ensured the safety of humans in the robot's operational environment.

Conclusions:

• The novel DRL framework effectively generates robot paths by learning from human demonstrations.
• The Koopman operator-based reward function accurately models human intent, improving learning efficiency and safety.
• The modified DQN algorithm shows promise for real-world robotic applications requiring human-robot collaboration.