The Impact of LiDAR Configuration on Goal-Based Navigation within a Deep Reinforcement Learning Framework.

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Technology

Background:

• Deep reinforcement learning (DRL) shows promise for autonomous robot navigation.
• Current DRL methods often use wide field of view (FOV) LiDAR sensors, which are expensive and unsuitable for small-scale applications.
• Optimizing sensor configuration is crucial for efficient DRL-based navigation.

Purpose of the Study:

• To investigate the impact of LiDAR sensor configuration on DRL model performance for mapless robot navigation.
• To propose a novel method for determining an optimal FOV for LiDAR sensors in DRL.
• To enhance collision avoidance and path planning capabilities in DRL models.

Main Methods:

• A novel approach to calculate an optimal FOV based on sensor width and minimum safe distance.
• Utilized LiDAR beams within the FOV, robot velocities, orientation, and distance to goal as DRL input state.
• Defined collision avoidance and path planning as the reward function for the DRL model.
• Trained and validated the DRL model with adjusted FOVs (±10°) to assess performance.

Main Results:

• The proposed LiDAR configuration with the computed FOV achieved a 98% success rate.
• The optimized DRL model demonstrated a lower time complexity of 0.25 m/s.
• Experimental validation on a Husky Robot confirmed the model's real-world applicability and performance.

Conclusions:

• The computed FOV significantly improves DRL performance for autonomous robot navigation and path planning.
• This method offers a cost-effective solution by optimizing existing LiDAR sensor configurations.
• The findings are applicable to real-world robotic systems requiring efficient collision avoidance and navigation.