Leveraging Expert Demonstration Features for Deep Reinforcement Learning in Floor Cleaning Robot Navigation

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Mobile cleaning robots require efficient navigation strategies.
• Learning optimal navigation from expert demonstrations is a promising approach.
• Deep Reinforcement Learning (DRL) offers a framework for complex control tasks.

Purpose of the Study:

• To develop a DRL-based approach for mobile cleaning robot navigation.
• To leverage expert demonstrations for learning robot motion commands.
• To optimize navigation policies and reduce training duration.

Main Methods:

• Collected expert demonstrations of robot motion trajectories in simulation.
• Extracted motion features: distance to obstacles and heading difference to goal.
• Optimized feature weights based on expert data.
• Created a reward function using feature values for semi-supervised DRL training.

Main Results:

• Demonstrated the viability of the DRL approach for robot navigation.
• Achieved reduced training time compared to traditional methods.
• Successfully trained a navigation policy that mimics expert behavior.

Conclusions:

• The proposed DRL method effectively learns mobile cleaning robot navigation commands.
• Expert demonstrations significantly enhance learning efficiency and policy quality.
• This approach offers a practical solution for developing intelligent robot navigation systems.