Table-Balancing Cooperative Robot Based on Deep Reinforcement Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Traditional reinforcement learning (RL) often focuses on single-agent tasks.
• Many real-world tasks, like table balancing, necessitate human-robot cooperation to ensure safety and efficiency.
• Existing research lacks robust methods for robots to dynamically adapt to human actions in collaborative tasks.

Purpose of the Study:

• To develop a deep reinforcement learning (DRL) based technique enabling robots to cooperatively balance a table with a human.
• To enable robots to recognize and respond to human behavior during a shared task.
• To enhance robot autonomy and safety in human-robot collaborative scenarios.

Main Methods:

• Utilized a deep Q-network (DQN), a DRL algorithm, for robot control.
• Integrated a camera system for the robot to perceive the table's state and human interaction.
• Trained the robot through simulated and real-world (H/W) experiments focusing on cooperative table balancing.

Main Results:

• The cooperative robot demonstrated an average optimal policy convergence rate of 90% across 20 training runs with optimized hyperparameters.
• In hardware experiments, the DQN-based robot achieved a 90% operational precision.
• The system successfully learned to balance the table in cooperation with a human, adapting to their movements.

Conclusions:

• The proposed DRL technique effectively enables robots to perform cooperative table balancing with humans.
• The DQN-based approach shows high performance and precision in recognizing human behavior and executing counter-balancing actions.
• This research validates the potential of DRL for developing sophisticated human-robot collaboration in dynamic physical tasks.