A fuzzy controller with supervised learning assisted reinforcement learning algorithm for obstacle avoidance.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Fuzzy logic systems offer efficient obstacle avoidance but struggle with human expert-based rule base maintenance.
• Reinforcement learning can automate fuzzy rule learning but faces challenges with extensive training and the curse of dimensionality.

Purpose of the Study:

• To propose a novel neural fuzzy system that overcomes limitations of traditional fuzzy logic and reinforcement learning for robot navigation.
• To develop an efficient two-phase learning approach combining supervised and reinforcement learning for robust obstacle avoidance.

Main Methods:

• A mixed coarse and fine learning strategy is employed: supervised learning for initial membership function determination, followed by reinforcement learning for fine-tuning.
• A modified Sutton and Barto reinforcement learning model is utilized to enhance exploration for more effective learning.
• A virtual environment (VE) simulator, offering desktop (DVE) and immersive (IVE) visualization, is developed to generate consistent training data.

Main Results:

• The proposed neural fuzzy system enables mobile robots to achieve collision-free navigation through a two-step tuning process.
• The virtual environment facilitates the acquisition of large, consistent datasets, mitigating challenges associated with real-world data collection.
• The enhanced exploration in the learning algorithm contributes to a more sufficiently learned and effective fuzzy rule base.

Conclusions:

• The developed neural fuzzy system provides an effective solution for autonomous mobile robot navigation and obstacle avoidance.
• The integration of supervised and reinforcement learning, coupled with a VE simulator, offers a scalable and efficient approach to training intelligent robotic systems.