Research on Reinforcement Learning-Based Autonomous Navigation and Obstacle Avoidance Methods for AGVs in Unknown...

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Reinforcement learning (RL) offers potential for autonomous navigation in hospital automated guided vehicles (AGVs).
• Challenges include designing reward functions, low sample efficiency, and slow convergence.
• Existing methods struggle with the complexities of unknown hospital environments.

Purpose of the Study:

• To propose BEAGM-PPO, a novel RL framework for autonomous navigation in hospital AGVs.
• To enhance sample efficiency and convergence speed in RL algorithms.
• To address the limitations of current autonomous navigation systems in healthcare settings.

Main Methods:

• Developed BEAGM-PPO, incorporating a reference model with expert demonstrations and policy derivation.
• Utilized behavior cloning and uncertainty estimation for policy derivation.
• Integrated ant colony optimization (ACO)-inspired pheromones and memory replay for improved exploration and action selection.

Main Results:

• BEAGM-PPO achieved the highest arrival rate in 3D simulation scenarios compared to baseline models.
• The imitation learning approach enabled uncertainty estimation for policy and model.
• Expanded training datasets further improved the performance of the autonomous navigation system.

Conclusions:

• BEAGM-PPO provides a robust framework for autonomous navigation in hospital AGVs.
• The method effectively improves sample efficiency and convergence behavior.
• This research lays a foundation for advanced autonomous systems in hospital logistics and patient care.