Quality-diversity based semi-autonomous teleoperation using reinforcement learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous systems achieve successes but often produce limited, repetitive behaviors.
• This lack of behavioral diversity can lead to inefficiencies, collisions, and safety concerns for human-robot interaction.
• Current robot control methods struggle to balance task performance with user-defined intentions.

Purpose of the Study:

• To develop a semi-autonomous teleoperation framework for enhanced robot controllability and diverse behaviors.
• To improve the efficiency and safety of robot operation through user-guided, high-level commands (options).
• To introduce a novel method for generating effective and varied robot behaviors.

Main Methods:

• A quality-diversity (QD) based sampling method is proposed to optimize both the quality and diversity of robot options.
• Reinforcement learning (RL) is utilized within the QD framework to learn optimal policies.
• A mixture of latent variable models is employed to learn multiple policy distributions representing distinct options.

Main Results:

• The proposed QD-based method demonstrates superior performance in simulation, achieving higher success rates and greater option diversity.
• Experiments show the framework effectively generates diverse and high-quality robot behaviors.
• Real-world tests indicate the method outperforms traditional manual keyboard control in cluttered environments.

Conclusions:

• The developed semi-autonomous teleoperation framework significantly enhances robot controllability and behavioral diversity.
• The QD-based approach offers an effective solution for generating varied robot behaviors, improving task efficiency and safety.
• This work advances robot learning by enabling more intuitive and adaptable human-robot interaction.