Mobile robots exploration through cnn-based reinforcement learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Mobile robot exploration is crucial for unknown environments.
• Traditional methods often require complex environment mapping or explicit programming.
• Deep learning advancements offer new possibilities for sensor-based navigation.

Purpose of the Study:

• To develop a reinforcement learning (RL) based exploration strategy for mobile robots.
• To enable robots to explore unknown corridor environments using only raw sensor data.
• To achieve autonomous obstacle avoidance and efficient exploration.

Main Methods:

• Utilized a deep Q-network (DQN) architecture for the RL agent.
• Employed a pre-trained convolutional neural network (CNN) to extract features from RGB-D sensor depth images.
• Trained the RL model in various simulated corridor environments.

Main Results:

• The robot controller demonstrated successful exploration capabilities in diverse simulated environments.
• The system achieved effective obstacle avoidance using only depth image input.
• The RL agent learned an exploration strategy directly from raw sensor information.

Conclusions:

• Reinforcement learning, combined with deep learning feature extraction, provides an effective method for mobile robot exploration.
• This approach eliminates the need for explicit environment models or pre-programmed navigation.
• It represents a significant advancement in autonomous robot navigation using raw sensory input.