A Reinforcement Learning Based Dirt-Exploration for Cleaning-Auditing Robot.

Area of Science:

• Robotics
• Artificial Intelligence
• Environmental Science

Background:

• Cleaning is a vital daily task, driving innovation in robotic cleaning solutions.
• Assessing cleaning quality is a critical, yet underexplored, research area.
• Defining
• How clean is clean
• necessitates effective auditing methods.

Purpose of the Study:

• To develop an autonomous cleaning-auditing robot capable of assessing area cleanliness.
• To propose a novel reinforcement learning-based strategy for dirt exploration.
• To address the challenge of "How clean is clean" through robotic auditing.

Main Methods:

• Utilized a reinforcement learning-based approach, specifically proximal policy approximation (PPO), for dirt exploration.
• Developed an experience-driven strategy to generate waypoints and sampling decisions for identifying dirt.
• Trained the policy network in simulated environments with diverse dirt patterns.

Main Results:

• The PPO-based strategy effectively guided the robot to explore probable dirt accumulation regions.
• Validated the trained policy in both simulated and real-world environments.
• Demonstrated the feasibility of the autonomous cleaning-auditing robot (BELUGA).

Conclusions:

• The proposed reinforcement learning strategy enables efficient dirt exploration for cleaning audits.
• Autonomous robots can effectively audit cleaning quality, addressing a key research gap.
• The BELUGA robot provides a practical platform for validating cleaning assessment technologies.