Enhancing Robot Navigation in Crowded Spaces Through Systematic Strategy Selection

This paper presents a framework for the online selection of autonomous navigation strategies in real-time for robots operating in crowded and dynamic environments. Traditional path-planning algorithms emphasize the computation of collision-free trajectories but may not account for more complex safety considerations that arise specifically when navigating around humans or other dynamic entities. To address these challenges, we propose a framework that evaluates a set of navigation strategies trained with deep reinforcement learning and selects the optimal strategy based on various metrics including safety, motion efficiency, and adaptability to changing scenarios. On one hand, the navigation strategies are trained using a reward function that incorporates objective metrics such as obstacle clearance and distance to the goal. On the other hand, the online evaluation of the strategies leverages a variety of additional performance metrics such as personal space intrusion, smoothness of path, control stability, and overall efficiency of control commands. The framework dynamically switches strategies to optimize navigation performance while balancing safety with control and path optimality. Simulated results show that the online selection of strategies can accomplish safer and more efficient navigation. Experimental results in real-world demonstrate the benefits of our approach, showing improved adaptability and context-aware decision-making compared to single-strategy navigation.