SOAR-RL: Safe and Open-Space Aware Reinforcement Learning for Mobile Robot Navigation in Narrow Spaces

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Autonomous navigation in narrow space environments (NSEs) is challenging due to dynamic pedestrian interactions and spatial constraints.
• Existing methods often rely on reactive collision avoidance, which is insufficient for proactive risk assessment in shared human-robot environments.

Purpose of the Study:

• To develop a deep reinforcement learning (DRL)-based navigation framework for mobile robots operating in human-robot shared environments.
• To enable robots to proactively identify and navigate safe paths while interacting with pedestrians in NSEs.

Main Methods:

• A DRL framework fusing 3D LiDAR and RGB camera data for real-time pedestrian detection, position, and velocity estimation.
• Construction of a human-aware occupancy map (HAOM) integrating static obstacles and dynamic risk zones as DRL input.
• Design of a novel state representation and reward structure to promote safe and proactive navigation behaviors.

Main Results:

• The proposed framework significantly improved navigation success rates in simulations of various NSE layouts (straight, L-shaped, cross-shaped).
• Collision incidents were substantially reduced compared to conventional navigation planners.
• The system demonstrated effective navigation in diverse dynamic obstacle scenarios.

Conclusions:

• The DRL-based navigation framework offers a robust solution for autonomous robot navigation in complex, dynamic narrow space environments.
• Integrating multi-sensor data and human-aware mapping enhances robot safety and efficiency in human-robot interactions.
• The approach overcomes limitations of traditional methods by enabling proactive risk assessment and safer route selection.